fdo-mirrors/NetworkManager
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/NetworkManager/NetworkManager.git synced 2026-05-14 09:38:04 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions
NetworkManager/src/libnm-client-impl/nm-client.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

9367 lines
346 KiB
C
Raw Normal View History

all: update deprecated SPDX license identifiers These SPDX license identifiers are deprecated ([1]). Update them. [1] https://spdx.org/licenses/ sed \ -e '1 s%^/\* SPDX-License-Identifier: \(GPL-2.0\|LGPL-2.1\)+ \*/$%/* SPDX-License-Identifier: \1-or-later */%' \ -e '1,2 s%^\(--\|#\|//\) SPDX-License-Identifier: \(GPL-2.0\|LGPL-2.1\)+$%\1 SPDX-License-Identifier: \2-or-later%' \ -i \ $(git grep -l SPDX-License-Identifier -- \ ':(exclude)shared/c-*/' \ ':(exclude)shared/n-*/' \ ':(exclude)shared/systemd/src' \ ':(exclude)src/systemd/src')
2020-12-23 22:21:36 +01:00
/* SPDX-License-Identifier: LGPL-2.1-or-later */
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/*
all: unify format of our Copyright source code comments ```bash readarray -d '' FILES < <( git ls-files -z \ ':(exclude)po' \ ':(exclude)shared/c-rbtree' \ ':(exclude)shared/c-list' \ ':(exclude)shared/c-siphash' \ ':(exclude)shared/c-stdaux' \ ':(exclude)shared/n-acd' \ ':(exclude)shared/n-dhcp4' \ ':(exclude)src/systemd/src' \ ':(exclude)shared/systemd/src' \ ':(exclude)m4' \ ':(exclude)COPYING*' ) sed \ -e 's/^\(--\|#\| \*\) *\(([cC]) *\)\?Copyright \+\(\(([cC])\) \+\)\?\(\(20\|19\)[0-9][0-9]\) *[-–] *\(\(20\|19\)[0-9][0-9]\) \+\([^ ].*\)$/\1 C1pyright#\5 - \7#\9/' \ -e 's/^\(--\|#\| \*\) *\(([cC]) *\)\?Copyright \+\(\(([cC])\) \+\)\?\(\(20\|19\)[0-9][0-9]\) *[,] *\(\(20\|19\)[0-9][0-9]\) \+\([^ ].*\)$/\1 C2pyright#\5, \7#\9/' \ -e 's/^\(--\|#\| \*\) *\(([cC]) *\)\?Copyright \+\(\(([cC])\) \+\)\?\(\(20\|19\)[0-9][0-9]\) \+\([^ ].*\)$/\1 C3pyright#\5#\7/' \ -e 's/^Copyright \(\(20\|19\)[0-9][0-9]\) \+\([^ ].*\)$/C4pyright#\1#\3/' \ -i \ "${FILES[@]}" echo ">>> untouched Copyright lines" git grep Copyright "${FILES[@]}" echo ">>> Copyright lines with unusual extra" git grep '\<C[0-9]pyright#' "${FILES[@]}" | grep -i reserved sed \ -e 's/\<C[0-9]pyright#\([^#]*\)#\(.*\)$/Copyright (C) \1 \2/' \ -i \ "${FILES[@]}" ``` https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/298
2019-10-01 09:20:35 +02:00
* Copyright (C) 2007 - 2008 Novell, Inc.
* Copyright (C) 2007 - 2018 Red Hat, Inc.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*/
build: move "libnm/" to "src/" and split it Like with "libnm-core/", split "libnm/" into different directories for the public headers, for the implementation and for the helper "aux" library.
2021-02-19 15:23:34 +01:00
#include "libnm-client-impl/nm-default-libnm.h"
all: consistently include config.h config.h should be included from every .c file, and it should be included before any other include. Fix that. (As a side effect of how I did this, this also changes us to consistently use "config.h" rather than <config.h>. To the extent that it matters [which is not much], quotes are more correct anyway, since we're talking about a file in our own build tree, not a system include.)
2014-11-13 10:07:02 -05:00
all: drop unnecessary includes of <errno.h> and <string.h> "nm-macros-interal.h" already includes <errno.h> and <string.h>. No need to include it everywhere else too.
2019-01-31 13:44:13 +01:00
#include "nm-client.h"
libnm/udev: cache and reuse udev instance for NMDevice No need to create a separate NMUdevClient instance for all devices. Instead, have one "struct udev" instance in NMClient and pass it down during object construction.
2017-03-20 12:27:17 +01:00
#include <libudev.h>
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
build: move "shared/nm-std-aux" to "src/libnm-std-aux"
2021-02-18 17:37:47 +01:00
#include "libnm-std-aux/c-list-util.h"
build: move "shared/nm-{glib-aux,log-null,log-core}" to "src/libnm-{glib-aux,log-null,log-core}"
2021-02-18 17:37:47 +01:00
#include "libnm-glib-aux/nm-c-list.h"
#include "libnm-glib-aux/nm-dbus-aux.h"
build: move "libnm-core/" to "src/" and split it "libnm-core/" is rather complicated. It provides a static library that is linked into libnm.so and NetworkManager. It also contains public headers (like "nm-setting.h") which are part of public libnm API. Then we have helper libraries ("libnm-core/nm-libnm-core-*/") which only rely on public API of libnm-core, but are themself static libraries that can be used by anybody who uses libnm-core. And "libnm-core/nm-libnm-core-intern" is used by libnm-core itself. Move "libnm-core/" to "src/". But also split it in different directories so that they have a clearer purpose. The goal is to have a flat directory hierarchy. The "src/libnm-core*/" directories correspond to the different modules (static libraries and set of headers that we have). We have different kinds of such modules because of how we combine various code together. The directory layout now reflects this.
2021-02-12 15:01:09 +01:00
#include "libnm-core-aux-intern/nm-common-macros.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-access-point.h"
#include "nm-active-connection.h"
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
#include "nm-checkpoint.h"
build: move "libnm-core/" to "src/" and split it "libnm-core/" is rather complicated. It provides a static library that is linked into libnm.so and NetworkManager. It also contains public headers (like "nm-setting.h") which are part of public libnm API. Then we have helper libraries ("libnm-core/nm-libnm-core-*/") which only rely on public API of libnm-core, but are themself static libraries that can be used by anybody who uses libnm-core. And "libnm-core/nm-libnm-core-intern" is used by libnm-core itself. Move "libnm-core/" to "src/". But also split it in different directories so that they have a clearer purpose. The goal is to have a flat directory hierarchy. The "src/libnm-core*/" directories correspond to the different modules (static libraries and set of headers that we have). We have different kinds of such modules because of how we combine various code together. The directory layout now reflects this.
2021-02-12 15:01:09 +01:00
#include "libnm-core-intern/nm-core-internal.h"
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
#include "nm-dbus-helpers.h"
libnm: replace nm-types.h by defining the types in respective headers The typedefs in nm-types.h confuse gtkdoc-scan. It generates a libnm-sections.txt file like this: <SECTION> <FILE>nm-types</FILE> <TITLE>NMDeviceOvs</TITLE> NMAccessPoint NMActiveConnection NMCheckpoint NMClient NMDevice ... Note the wrongly picked title and, more importantly, the object types in a bogus section. This in turn makes gtkdoc-mkdb fail to include the property and signal documentation in appropriate sections. Without nm-types.h, we need to mind the header dependencies. This means that we need to order the headers that define types before the ones that use them. Also, we need to break the depencency loops in few palces.
2022-05-05 15:39:45 +02:00
#include "nm-wifi-p2p-peer.h"
libnm: add support form 6LoWPAN devices
2018-05-22 16:45:05 +02:00
#include "nm-device-6lowpan.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-device-adsl.h"
#include "nm-device-bond.h"
#include "nm-device-bridge.h"
#include "nm-device-bt.h"
core: add support for dummy devices Add support for creating dummy devices. This commit adds a D-Bus interface 'org.freedesktop.NetworkManager.Device.Dummy' which is used primarily for determining the device type but does not carry any properties.
2017-01-31 14:14:33 +01:00
#include "nm-device-dummy.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-device-ethernet.h"
#include "nm-device-generic.h"
geneve: added GENEVE device support Support device type geneve in libnm and nmcli. (cherry picked from commit 0bfb8fa89d4ef5afd846f9893197ffc6a2e7b20a)
2026-02-02 19:02:29 -05:00
#include "nm-device-geneve.h"
HSR: add support to HSR/PRP interface This patch add support to HSR/PRP interface. Please notice that PRP driver is represented as HSR too. They are different drivers but on kernel they are integrated together. HSR/PRP is a network protocol standard for Ethernet that provides seamless failover against failure of any network component. It intends to be transparent to the application. These protocols are useful for applications that request high availability and short switchover time e.g electrical substation or high power inverters. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1791
2023-10-27 14:01:13 +02:00
#include "nm-device-hsr.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-device-infiniband.h"
#include "nm-device-ip-tunnel.h"
ipvlan: add support to IPVLAN interface This patch add support to IPVLAN interface. IPVLAN is a driver for a virtual network device that can be used in container environment to access the host network. IPVLAN exposes a single MAC address to the external network regardless the number of IPVLAN device created inside the host network. This means that a user can have multiple IPVLAN devices in multiple containers and the corresponding switch reads a single MAC address. IPVLAN driver is useful when the local switch imposes constraints on the total number of MAC addresses that it can manage.
2024-09-06 14:14:24 +02:00
#include "nm-device-ipvlan.h"
support loopback interface Support managing the loopback interface through NM as the users want to set the proper mtu for loopback interface when forwarding the packets. Additionally, the IP addresses, DNS, route and routing rules are also allowed to configure for the loopback connection profiles. https://bugzilla.redhat.com/show_bug.cgi?id=2060905
2022-06-12 19:50:09 -04:00
#include "nm-device-loopback.h"
core,libnm: introduce NMDeviceMacsec At the moment the device only exposes the current link status, but cannot create new links.
2016-06-30 18:20:22 +02:00
#include "nm-device-macsec.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-device-macvlan.h"
#include "nm-device-modem.h"
wifi-p2p: rename files for consistent Wi-Fi P2P naming We named the types inconsistently: - "p2p-wireless" ("libnm-core/nm-setting-p2p-wireless.h") - "p2p" ("libnm/nm-p2p-peer.h") - "p2p-wifi" ("src/devices/wifi/nm-device-p2p-wifi.h") It seems to me, "libnm/nm-p2p-peer.h" should be qualified with a "Wi-Fi" specific name. It's not just peer-to-peer, it's Wi-Fi P2P. Yes, there is an inconsistency now, because there is already "libnm/nm-access-point.h". It seems to me (from looking at the internet), that the name "Wi-Fi P2P" is more common than "P2P Wi-Fi" -- although both are used. There is also the name "Wi-Fi Direct". But it's not clear which name should be preferred here, so stick to "Wi-Fi P2P". In this first commit only rename the files. The following commit will rename the content.
2019-01-28 12:39:38 +01:00
#include "nm-device-ovs-bridge.h"
libnm: add support for ovs-interface devices
2017-10-10 11:04:32 +02:00
#include "nm-device-ovs-interface.h"
libnm: add support for ovs-port devices
2017-10-10 11:04:32 +02:00
#include "nm-device-ovs-port.h"
device: add NMDevicePPP The new device type represents a PPP interface, and will implement the activation of new-style PPPoE connections, i.e. the ones that don't claim the parent device.
2017-06-06 15:55:08 +02:00
#include "nm-device-ppp.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-device-team.h"
#include "nm-device-tun.h"
#include "nm-device-vlan.h"
#include "nm-device-vxlan.h"
wifi-p2p: rename files for consistent Wi-Fi P2P naming We named the types inconsistently: - "p2p-wireless" ("libnm-core/nm-setting-p2p-wireless.h") - "p2p" ("libnm/nm-p2p-peer.h") - "p2p-wifi" ("src/devices/wifi/nm-device-p2p-wifi.h") It seems to me, "libnm/nm-p2p-peer.h" should be qualified with a "Wi-Fi" specific name. It's not just peer-to-peer, it's Wi-Fi P2P. Yes, there is an inconsistency now, because there is already "libnm/nm-access-point.h". It seems to me (from looking at the internet), that the name "Wi-Fi P2P" is more common than "P2P Wi-Fi" -- although both are used. There is also the name "Wi-Fi Direct". But it's not clear which name should be preferred here, so stick to "Wi-Fi P2P". In this first commit only rename the files. The following commit will rename the content.
2019-01-28 12:39:38 +01:00
#include "nm-device-wifi-p2p.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-device-wifi.h"
libnm: introduce NMDeviceWireGuard
2018-03-13 13:42:38 +00:00
#include "nm-device-wireguard.h"
libnm: add support for WPAN devices
2018-03-09 17:19:36 +01:00
#include "nm-device-wpan.h"
libnm: replace nm-types.h by defining the types in respective headers The typedefs in nm-types.h confuse gtkdoc-scan. It generates a libnm-sections.txt file like this: <SECTION> <FILE>nm-types</FILE> <TITLE>NMDeviceOvs</TITLE> NMAccessPoint NMActiveConnection NMCheckpoint NMClient NMDevice ... Note the wrongly picked title and, more importantly, the object types in a bogus section. This in turn makes gtkdoc-mkdb fail to include the property and signal documentation in appropriate sections. Without nm-types.h, we need to mind the header dependencies. This means that we need to order the headers that define types before the ones that use them. Also, we need to break the depencency loops in few palces.
2022-05-05 15:39:45 +02:00
#include "nm-device-olpc-mesh.h"
wifi-p2p: rename files for consistent Wi-Fi P2P naming We named the types inconsistently: - "p2p-wireless" ("libnm-core/nm-setting-p2p-wireless.h") - "p2p" ("libnm/nm-p2p-peer.h") - "p2p-wifi" ("src/devices/wifi/nm-device-p2p-wifi.h") It seems to me, "libnm/nm-p2p-peer.h" should be qualified with a "Wi-Fi" specific name. It's not just peer-to-peer, it's Wi-Fi P2P. Yes, there is an inconsistency now, because there is already "libnm/nm-access-point.h". It seems to me (from looking at the internet), that the name "Wi-Fi P2P" is more common than "P2P Wi-Fi" -- although both are used. There is also the name "Wi-Fi Direct". But it's not clear which name should be preferred here, so stick to "Wi-Fi P2P". In this first commit only rename the files. The following commit will rename the content.
2019-01-28 12:39:38 +01:00
#include "nm-dhcp-config.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-dhcp4-config.h"
#include "nm-dhcp6-config.h"
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
#include "nm-dns-manager.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-ip4-config.h"
#include "nm-ip6-config.h"
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
#include "nm-object-private.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-remote-connection.h"
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
#include "nm-utils.h"
libnm: add ethtool fec backported symbols from 1.50.4 Add to main branch ethtool fec symbols that were backported to 1.50.4 to allow seamless upgrading from 1.50 to 1.54.
2025-04-27 22:30:16 -04:00
#include "nm-setting-ethtool.h"
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
#include "nm-vpn-connection.h"
libnm: port to GDBus Port libnm-core/libnm to GDBus. The NetworkManager daemon continues to use dbus-glib; the previously-added connection hash/variant conversion methods are now moved to NetworkManagerUtils (along with a few other utilities that are now only needed by the daemon code).
2014-09-10 13:51:53 -04:00
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
/*****************************************************************************/
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
all: require a semicolon after NM_CACHED_QUARK_FCN()
2020-09-23 10:53:06 +02:00
NM_CACHED_QUARK_FCN("nm-context-busy-watcher", nm_context_busy_watcher_quark);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
static void
_context_busy_watcher_attach_integration_source_cb(gpointer data, GObject *where_the_object_was)
{
nm_g_source_destroy_and_unref(data);
}
libnm: expose nm_context_busy_watcher_integrate_source() as internal API for reuse This will also be useful for NMSecretAgentOld. The mechanics how NMClient handles the GMainContext and the context-busy-watcher apply really to every GObject that uses GDBusConnection and registers to signals. At least, as long as the API provides no shutdown/stop method, because that means shutdown/stop happens when unreferencing the instance, at which point pending operations get cancelled (but they cannot complete right away due to the nature of GTask and g_dbus_connection_call()). If there is a shutdown/stop API, then all pending operations could keep the instance alive, and the instance would sticks around (and keeps the GMainContext busy) until shutdown is completed. Basically, then the instance could be the context-busy-watcher itself. But in existing API which does not require the user to explicitly shutdown, that is not a feasible (backward compatible) addition. But the context-busy-watcher object is.
2019-12-17 13:59:15 +01:00
void
nm_context_busy_watcher_integrate_source(GMainContext *outer_context,
GMainContext *inner_context,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GObject *context_busy_watcher)
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
{
libnm: expose nm_context_busy_watcher_integrate_source() as internal API for reuse This will also be useful for NMSecretAgentOld. The mechanics how NMClient handles the GMainContext and the context-busy-watcher apply really to every GObject that uses GDBusConnection and registers to signals. At least, as long as the API provides no shutdown/stop method, because that means shutdown/stop happens when unreferencing the instance, at which point pending operations get cancelled (but they cannot complete right away due to the nature of GTask and g_dbus_connection_call()). If there is a shutdown/stop API, then all pending operations could keep the instance alive, and the instance would sticks around (and keeps the GMainContext busy) until shutdown is completed. Basically, then the instance could be the context-busy-watcher itself. But in existing API which does not require the user to explicitly shutdown, that is not a feasible (backward compatible) addition. But the context-busy-watcher object is.
2019-12-17 13:59:15 +01:00
GSource *source;
nm_assert(outer_context);
nm_assert(inner_context);
nm_assert(outer_context != inner_context);
nm_assert(G_IS_OBJECT(context_busy_watcher));
source = nm_utils_g_main_context_create_integrate_source(inner_context);
g_source_attach(source, outer_context);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
/* The problem is...
*
* NMClient is associated with a GMainContext, just like its underlying GDBusConnection
* also queues signals and callbacks on that main context. During operation, NMClient
libnm/doc: clarify GMainContext handling in NMClient documentation https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/819
2021-04-18 08:34:27 +02:00
* will schedule async operations which will return asynchronously on that GMainContext.
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
*
* Note that depending on whether NMClient got initialized synchronously or asynchronously,
* it has an internal priv->dbus_context that is different from the outer priv->main_context.
* However, the problem is in both cases.
*
* So, as long as there are pending D-Bus calls, the GMainContext is referenced and kept alive.
* When NMClient gets destroyed, the pending calls get cancelled, but the async callback are still
* scheduled to return.
* That means, the main context stays alive until it gets iterated long enough so that all pending
* operations are completed.
*
* Note that pending operations don't keep NMClient alive, so NMClient can already be gone by
* then, but the user still should iterate the main context long enough to process the (cancelled)
* callbacks... at least, if the user cares about whether the remaining memory and file descriptors
* of the GMainContext can be reclaimed.
*
* In hindsight, maybe pending references should kept NMClient alive. But then NMClient would
* need a special "shutdown()" API that the user must invoke, because unrefing would no longer
* be enough to ensure a shutdown (imagine a situation where NMClient receives a constant flow
* of "CheckPermissions" signals, which keeps retriggering an async request). Anyway, we cannot
* add such a shutdown API now, as it would break client's expectations that they can just unref
* the NMClient to destroy it.
*
* So, we allow NMClient to unref, but the user is advised to keep iterating the main context.
* But for how long? Here comes nm_client_get_context_busy_watcher() into play. The user may
* subscribe a weak pointer to that instance and should keep iterating as long as the object
* exists.
*
libnm/doc: clarify GMainContext handling in NMClient documentation https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/819
2021-04-18 08:34:27 +02:00
* Now, back to synchronous initialization: here we have the internal priv->dbus_context context.
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
* We also cannot remove that context right away, instead we need to keep it integrated in the
* caller's priv->main_context as long as we have pending calls: that is, as long as the
* context-busy-watcher is alive.
*/
g_object_weak_ref(context_busy_watcher,
_context_busy_watcher_attach_integration_source_cb,
libnm: expose nm_context_busy_watcher_integrate_source() as internal API for reuse This will also be useful for NMSecretAgentOld. The mechanics how NMClient handles the GMainContext and the context-busy-watcher apply really to every GObject that uses GDBusConnection and registers to signals. At least, as long as the API provides no shutdown/stop method, because that means shutdown/stop happens when unreferencing the instance, at which point pending operations get cancelled (but they cannot complete right away due to the nature of GTask and g_dbus_connection_call()). If there is a shutdown/stop API, then all pending operations could keep the instance alive, and the instance would sticks around (and keeps the GMainContext busy) until shutdown is completed. Basically, then the instance could be the context-busy-watcher itself. But in existing API which does not require the user to explicitly shutdown, that is not a feasible (backward compatible) addition. But the context-busy-watcher object is.
2019-12-17 13:59:15 +01:00
source);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
}
/*****************************************************************************/
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
typedef struct {
/* It is quite wasteful to require 2 pointers per property (of an instance) only to track whether
* the property got changed. But it's convenient! */
CList changed_prop_lst;
GVariant *prop_data_value;
} NMLDBusObjPropData;
typedef struct {
CList iface_lst;
union {
const NMLDBusMetaIface *meta;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMRefString *name;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
} dbus_iface;
CList changed_prop_lst_head;
/* We also keep track of non-well known interfaces. The presence of a D-Bus interface
* is what makes a D-Bus alive or not. As we should track all D-Bus objects, we also
* need to track whether there are any interfaces on it -- even if we otherwise don't
* care about the interface. */
bool dbus_iface_is_wellknown : 1;
/* if TRUE, the interface is about to be removed. */
bool iface_removed : 1;
bool nmobj_checked : 1;
bool nmobj_compatible : 1;
NMLDBusObjPropData prop_datas[];
} NMLDBusObjIfaceData;
/* The dbus_path must be the first element, so when we hash the object by the dbus_path,
* we also can lookup the object by only having a NMRefString at hand
* using nm_pdirect_hash()/nm_pdirect_equal(). */
G_STATIC_ASSERT(G_STRUCT_OFFSET(NMLDBusObject, dbus_path) == 0);
typedef void (*NMLDBusObjWatchNotifyFcn)(NMClient *client, gpointer obj_watcher);
struct _NMLDBusObjWatcher {
NMLDBusObject *dbobj;
struct {
CList watcher_lst;
NMLDBusObjWatchNotifyFcn notify_fcn;
} _priv;
};
typedef struct {
NMLDBusObjWatcher parent;
gpointer user_data;
} NMLDBusObjWatcherWithPtr;
/*****************************************************************************/
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
NM_GOBJECT_PROPERTIES_DEFINE(NMClient,
libnm: add NM_CLIENT_DBUS_CONNECTION property The used GDBusConnection should be configurable when creating the NMClient instance. Automatically choosing one of the g_bus_get() singletons is fine by default, but it's an unnecessary limitation.
2019-10-12 14:59:23 +02:00
PROP_DBUS_CONNECTION,
libnm: add NM_CLIENT_DBUS_NAME_OWNER property It's not yet implemented. But obviously it's interesting to get the name owner to which the NMClient is currently connected. Note only that: the name-owner property really says whether NM is currently running or not.
2019-10-15 16:08:36 +02:00
PROP_DBUS_NAME_OWNER,
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
PROP_VERSION,
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
PROP_INSTANCE_FLAGS,
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
PROP_STATE,
PROP_STARTUP,
libnm: synchronize NMClient and NMRemoteSettings "is NM running" properties Rename NMClient:manager-running and NMRemoteSettings:service-running to both be :nm-running.
2014-08-05 10:05:26 -04:00
PROP_NM_RUNNING,
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
PROP_NETWORKING_ENABLED,
PROP_WIRELESS_ENABLED,
PROP_WIRELESS_HARDWARE_ENABLED,
PROP_WWAN_ENABLED,
PROP_WWAN_HARDWARE_ENABLED,
PROP_WIMAX_ENABLED,
PROP_WIMAX_HARDWARE_ENABLED,
libnm: support radio flags
2022-03-21 10:20:11 +01:00
PROP_RADIO_FLAGS,
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
PROP_ACTIVE_CONNECTIONS,
PROP_CONNECTIVITY,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
PROP_CONNECTIVITY_CHECK_URI,
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
PROP_CONNECTIVITY_CHECK_AVAILABLE,
PROP_CONNECTIVITY_CHECK_ENABLED,
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
PROP_PRIMARY_CONNECTION,
PROP_ACTIVATING_CONNECTION,
PROP_DEVICES,
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
PROP_ALL_DEVICES,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
PROP_CONNECTIONS,
PROP_HOSTNAME,
PROP_CAN_MODIFY,
nm-manager: add 'metered' property This introduces a global metered property which makes easier for clients to obtain the metered status of the current primary connection.
2015-06-03 09:15:24 +02:00
PROP_METERED,
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
PROP_DNS_MODE,
PROP_DNS_RC_MANAGER,
PROP_DNS_CONFIGURATION,
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
PROP_CHECKPOINTS,
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
PROP_VERSION_INFO,
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
PROP_CAPABILITIES,
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
PROP_PERMISSIONS_STATE, );
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
enum {
DEVICE_ADDED,
DEVICE_REMOVED,
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
ANY_DEVICE_ADDED,
ANY_DEVICE_REMOVED,
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
PERMISSION_CHANGED,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
CONNECTION_ADDED,
CONNECTION_REMOVED,
libnm/client: emit signals when active connection disappears It allows us to reliably track failures to activate a connection.
2016-07-07 11:35:19 +02:00
ACTIVE_CONNECTION_ADDED,
ACTIVE_CONNECTION_REMOVED,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
LAST_SIGNAL
};
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static guint signals[LAST_SIGNAL] = {0};
enum {
PROPERTY_O_IDX_NM_ACTIVATING_CONNECTION = 0,
PROPERTY_O_IDX_NM_PRIMAY_CONNECTION,
_PROPERTY_O_IDX_NM_NUM,
};
enum {
PROPERTY_AO_IDX_DEVICES = 0,
PROPERTY_AO_IDX_ALL_DEVICES,
PROPERTY_AO_IDX_ACTIVE_CONNECTIONS,
PROPERTY_AO_IDX_CHECKPOINTS,
_PROPERTY_AO_IDX_NM_NUM,
};
typedef struct {
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
struct udev *udev;
GMainContext *main_context;
GMainContext *dbus_context;
GObject *context_busy_watcher;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GDBusConnection *dbus_connection;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLInitData *init_data;
GHashTable *dbus_objects;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
CList obj_changed_lst_head;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *name_owner_get_cancellable;
GCancellable *get_managed_objects_cancellable;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
CList queue_notify_lst_head;
CList notify_event_lst_head;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
CList dbus_objects_lst_head_watched_only;
CList dbus_objects_lst_head_on_dbus;
CList dbus_objects_lst_head_with_nmobj_not_ready;
CList dbus_objects_lst_head_with_nmobj_ready;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObject *dbobj_nm;
NMLDBusObject *dbobj_settings;
NMLDBusObject *dbobj_dns_manager;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
gsize log_call_counter;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
guint8 *permissions;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GCancellable *permissions_cancellable;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
char *name_owner;
guint name_owner_changed_id;
guint dbsid_nm_object_manager;
guint dbsid_dbus_properties_properties_changed;
guint dbsid_nm_settings_connection_updated;
guint dbsid_nm_connection_active_state_changed;
guint dbsid_nm_vpn_connection_state_changed;
guint dbsid_nm_check_permissions;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add "initialized-{good,bad}" flags to NMClientInstanceFlags When using async initialization with GAsyncInitable, the user usually can only know that initialization is complete by passing a callback. In simple cases, that can be cumbersome. Also expose a flag that allows to poll that information. Reuse the existing NM_CLIENT_INSTANCE_FLAGS for that. There is an ugliness here, that suddenly there are instance flags that cannot be set, but are still returned by the getter. But as this is a relatively obscure feature, it seems more lightweight to implement it this way (instead of adding a separate property and getter function).
2022-10-06 10:13:14 +02:00
NMClientInstanceFlags instance_flags : 5;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
NMTernary permissions_state : 3;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
bool instance_flags_constructed : 1;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
bool udev_inited : 1;
bool notify_event_lst_changed : 1;
bool check_dbobj_visible_all : 1;
bool nm_running : 1;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
struct {
NMLDBusPropertyO property_o[_PROPERTY_O_IDX_NM_NUM];
NMLDBusPropertyAO property_ao[_PROPERTY_AO_IDX_NM_NUM];
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
char *connectivity_check_uri;
char *version;
guint32 *capabilities_arr;
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
guint32 *version_info_arr;
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
gsize capabilities_len;
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
gsize version_info_len;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
guint32 connectivity;
guint32 state;
guint32 metered;
libnm: support radio flags
2022-03-21 10:20:11 +01:00
guint32 radio_flags;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
bool connectivity_check_available;
bool connectivity_check_enabled;
bool networking_enabled;
bool startup;
bool wireless_enabled;
bool wireless_hardware_enabled;
bool wwan_enabled;
bool wwan_hardware_enabled;
} nm;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
struct {
NMLDBusPropertyAO connections;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
char *hostname;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
bool can_modify;
} settings;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
struct {
GPtrArray *configuration;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
char *mode;
char *rc_manager;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
} dns_manager;
} NMClientPrivate;
struct _NMClient {
union {
GObject parent;
NMObjectBase obj_base;
};
NMClientPrivate _priv;
};
struct _NMClientClass {
union {
GObjectClass parent;
NMObjectBaseClass obj_base;
};
};
static void nm_client_initable_iface_init(GInitableIface *iface);
static void nm_client_async_initable_iface_init(GAsyncInitableIface *iface);
G_DEFINE_TYPE_WITH_CODE(NMClient,
nm_client,
G_TYPE_OBJECT,
G_IMPLEMENT_INTERFACE(G_TYPE_INITABLE, nm_client_initable_iface_init);
G_IMPLEMENT_INTERFACE(G_TYPE_ASYNC_INITABLE,
nm_client_async_initable_iface_init);)
#define NM_CLIENT_GET_PRIVATE(self) _NM_GET_PRIVATE(self, NMClient, NM_IS_CLIENT)
/*****************************************************************************/
static void _init_start_check_complete(NMClient *self);
static void name_owner_changed_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *object_path,
const char *interface_name,
const char *signal_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer user_data);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
static void name_owner_get_call(NMClient *self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
static void _set_nm_running(NMClient *self, gboolean queue_notify);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
static NMRefString *_dbus_path_nm = NULL;
static NMRefString *_dbus_path_settings = NULL;
static NMRefString *_dbus_path_dns_manager = NULL;
/*****************************************************************************/
shared: drop _STATIC variant of macros that define functions Several macros are used to define function. They had a "_STATIC" variant, to define the function as static. I think those macros should not try to abstract entirely what they do. They should not accept the function scope as argument (or have two variants per scope). This also because it might make sense to add additional __attribute__(()) to the function. That only works, if the macro does not pretend to *not* define a plain function. Instead, embrace what the function does and let the users place the function scope as they see fit. This also follows what is already done with static NM_CACHED_QUARK_FCN ("autoconnect-root", autoconnect_root_quark)
2020-02-13 14:55:26 +01:00
static NM_UTILS_LOOKUP_STR_DEFINE(
nml_dbus_obj_state_to_string,
NMLDBusObjState,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NM_UTILS_LOOKUP_DEFAULT_WARN("???"),
NM_UTILS_LOOKUP_ITEM(NML_DBUS_OBJ_STATE_UNLINKED, "unlinked"),
NM_UTILS_LOOKUP_ITEM(NML_DBUS_OBJ_STATE_WATCHED_ONLY, "watched-only"),
NM_UTILS_LOOKUP_ITEM(NML_DBUS_OBJ_STATE_ON_DBUS, "on-dbus"),
NM_UTILS_LOOKUP_ITEM(NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY, "not-ready"),
NM_UTILS_LOOKUP_ITEM(NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY, "ready"), );
/*****************************************************************************/
/**
* nm_client_error_quark:
*
* Registers an error quark for #NMClient if necessary.
*
* Returns: the error quark used for #NMClient errors.
**/
all: require a semicolon after NM_CACHED_QUARK_FCN()
2020-09-23 10:53:06 +02:00
NM_CACHED_QUARK_FCN("nm-client-error-quark", nm_client_error_quark);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
NMLInitData *
nml_init_data_new_sync(GCancellable *cancellable, GMainLoop *main_loop, GError **error_location)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
NMLInitData *init_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
init_data = g_slice_new(NMLInitData);
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
*init_data = (NMLInitData) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
.cancellable = nm_g_object_ref(cancellable),
.is_sync = TRUE,
.data.sync =
{
.main_loop = main_loop,
.error_location = error_location,
},
};
return init_data;
}
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
NMLInitData *
nml_init_data_new_async(GCancellable *cancellable, GTask *task_take)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
NMLInitData *init_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
init_data = g_slice_new(NMLInitData);
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
*init_data = (NMLInitData) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
.cancellable = nm_g_object_ref(cancellable),
.is_sync = FALSE,
.data.async =
{
.task = g_steal_pointer(&task_take),
},
};
return init_data;
}
libnm: factor out nml_init_data_return() for reuse
2019-12-18 17:12:45 +01:00
void
nml_init_data_return(NMLInitData *init_data, GError *error_take)
{
nm_assert(init_data);
nm_clear_pointer(&init_data->cancel_on_idle_source, nm_g_source_destroy_and_unref);
nm_clear_g_signal_handler(init_data->cancellable, &init_data->cancelled_id);
if (init_data->is_sync) {
if (error_take)
g_propagate_error(init_data->data.sync.error_location, error_take);
g_main_loop_quit(init_data->data.sync.main_loop);
} else {
if (error_take)
g_task_return_error(init_data->data.async.task, error_take);
else
g_task_return_boolean(init_data->data.async.task, TRUE);
g_object_unref(init_data->data.async.task);
}
nm_g_object_unref(init_data->cancellable);
nm_g_slice_free(init_data);
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
GError *
_nm_client_new_error_nm_not_running(void)
{
return g_error_new_literal(NM_CLIENT_ERROR,
NM_CLIENT_ERROR_MANAGER_NOT_RUNNING,
"NetworkManager is not running");
}
GError *
_nm_client_new_error_nm_not_cached(void)
{
return g_error_new_literal(NM_CLIENT_ERROR,
NM_CLIENT_ERROR_FAILED,
"Object is no longer in the client cache");
}
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
static void
all/trivial: rename local variable for user_data for nm_utils_user_data_unpack() In almost all cases, the variable of this kind is named "user_data". Rename it for consistency.
2020-10-22 15:07:36 +02:00
_nm_client_dbus_call_simple_cb(GObject *source, GAsyncResult *result, gpointer user_data)
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GAsyncReadyCallback callback;
gpointer callback_user_data;
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
gs_unref_object GObject *context_busy_watcher = NULL;
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
gpointer obfuscated_self_ptr;
gpointer log_call_counter_ptr;
all/trivial: rename local variable for user_data for nm_utils_user_data_unpack() In almost all cases, the variable of this kind is named "user_data". Rename it for consistency.
2020-10-22 15:07:36 +02:00
nm_utils_user_data_unpack(user_data,
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
&callback,
all/trivial: rename local variable for user_data for nm_utils_user_data_unpack() In almost all cases, the variable of this kind is named "user_data". Rename it for consistency.
2020-10-22 15:07:36 +02:00
&callback_user_data,
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
&context_busy_watcher,
&obfuscated_self_ptr,
&log_call_counter_ptr);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
NML_DBUS_LOG(_NML_NMCLIENT_LOG_LEVEL_COERCE(NML_DBUS_LOG_LEVEL_TRACE),
"nmclient[" NM_HASH_OBFUSCATE_PTR_FMT "]: call[%" G_GSIZE_FORMAT "] completed",
(guint64) GPOINTER_TO_SIZE(obfuscated_self_ptr),
GPOINTER_TO_SIZE(log_call_counter_ptr));
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
all/trivial: rename local variable for user_data for nm_utils_user_data_unpack() In almost all cases, the variable of this kind is named "user_data". Rename it for consistency.
2020-10-22 15:07:36 +02:00
callback(source, result, callback_user_data);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_client_dbus_call_simple(NMClient *self,
GCancellable *cancellable,
const char *object_path,
const char *interface_name,
const char *method_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const GVariantType *reply_type,
GDBusCallFlags flags,
int timeout_msec,
GAsyncReadyCallback callback,
gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_auto_pop_gmaincontext GMainContext *dbus_context = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_free char *log_str = NULL;
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
gsize log_call_counter;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(priv->name_owner);
nm_assert(!cancellable || G_IS_CANCELLABLE(cancellable));
nm_assert(callback);
nm_assert(object_path);
nm_assert(interface_name);
nm_assert(method_name);
nm_assert(reply_type);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_context = nm_g_main_context_push_thread_default_if_necessary(priv->dbus_context);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
log_call_counter = ++priv->log_call_counter;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add trace logging for D-Bus calls libnm-dbus: <trace> [394618.09943] nmclient[97de7f355f14dfa9]: call D-Bus method on :1.17: /org/freedesktop/NetworkManager/Settings/17, org.freedesktop.NetworkManager.Settings.Connection.GetSettings -> (a{sa{sv}}) (())
2020-07-15 11:16:30 +02:00
NML_NMCLIENT_LOG_T(self,
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
"call[%" G_GSIZE_FORMAT "] D-Bus method on %s: %s, %s.%s -> %s (%s)",
log_call_counter,
libnm: add trace logging for D-Bus calls libnm-dbus: <trace> [394618.09943] nmclient[97de7f355f14dfa9]: call D-Bus method on :1.17: /org/freedesktop/NetworkManager/Settings/17, org.freedesktop.NetworkManager.Settings.Connection.GetSettings -> (a{sa{sv}}) (())
2020-07-15 11:16:30 +02:00
priv->name_owner,
object_path,
interface_name,
method_name,
(const char *) reply_type ?: "???",
parameters ? (log_str = g_variant_print(parameters, TRUE)) : "NULL");
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_dbus_connection_call(priv->dbus_connection,
priv->name_owner,
object_path,
interface_name,
method_name,
parameters,
reply_type,
flags,
timeout_msec,
cancellable,
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
_nm_client_dbus_call_simple_cb,
libnm: trace log the call and the completion event for D-Bus methods
2020-07-15 12:39:50 +02:00
nm_utils_user_data_pack(callback,
user_data,
g_object_ref(priv->context_busy_watcher),
GSIZE_TO_POINTER(NM_HASH_OBFUSCATE_PTR(self)),
GSIZE_TO_POINTER(log_call_counter)));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_client_dbus_call(NMClient *self,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer source_obj,
gpointer source_tag,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GAsyncReadyCallback user_callback,
gpointer user_callback_data,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *object_path,
const char *interface_name,
const char *method_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const GVariantType *reply_type,
GDBusCallFlags flags,
int timeout_msec,
GAsyncReadyCallback internal_callback)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_object GTask *task = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(!source_obj || G_IS_OBJECT(source_obj));
nm_assert(source_tag);
nm_assert(!cancellable || G_IS_CANCELLABLE(cancellable));
nm_assert(internal_callback);
nm_assert(object_path);
nm_assert(interface_name);
nm_assert(method_name);
nm_assert(reply_type);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
task = nm_g_task_new(source_obj, cancellable, source_tag, user_callback, user_callback_data);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!self) {
libnm: fix "parameters" argument in nm_client_dbus_call() to be optional It was documented to be an optional parameter. That is also in line with g_dbus_connection_call(), which is essentially wrapped by nm_client_dbus_call(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2022-07-04 13:46:39 +02:00
if (parameters)
nm_g_variant_unref_floating(parameters);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_task_return_error(task, _nm_client_new_error_nm_not_cached());
return;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv = NM_CLIENT_GET_PRIVATE(self);
if (!priv->name_owner) {
libnm: fix "parameters" argument in nm_client_dbus_call() to be optional It was documented to be an optional parameter. That is also in line with g_dbus_connection_call(), which is essentially wrapped by nm_client_dbus_call(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2022-07-04 13:46:39 +02:00
if (parameters)
nm_g_variant_unref_floating(parameters);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_task_return_error(task, _nm_client_new_error_nm_not_running());
return;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call_simple(self,
cancellable,
object_path,
interface_name,
method_name,
parameters,
reply_type,
flags,
timeout_msec,
internal_callback,
g_steal_pointer(&task));
}
GVariant *
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_client_dbus_call_sync(NMClient *self,
GCancellable *cancellable,
const char *object_path,
const char *interface_name,
const char *method_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const GVariantType *reply_type,
GDBusCallFlags flags,
int timeout_msec,
gboolean strip_dbus_error,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GError **error)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_variant GVariant *ret = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(!cancellable || G_IS_CANCELLABLE(cancellable));
nm_assert(!error || !*error);
nm_assert(object_path);
nm_assert(interface_name);
nm_assert(method_name);
nm_assert(parameters);
nm_assert(reply_type);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!self) {
nm_g_variant_unref_floating(parameters);
nm_g_set_error_take_lazy(error, _nm_client_new_error_nm_not_cached());
return NULL;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv = NM_CLIENT_GET_PRIVATE(self);
if (!priv->name_owner) {
nm_g_variant_unref_floating(parameters);
nm_g_set_error_take_lazy(error, _nm_client_new_error_nm_not_running());
return NULL;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = g_dbus_connection_call_sync(priv->dbus_connection,
priv->name_owner,
object_path,
interface_name,
method_name,
parameters,
reply_type,
flags,
timeout_msec,
cancellable,
error);
if (!ret) {
if (error && strip_dbus_error)
g_dbus_error_strip_remote_error(*error);
return NULL;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return g_steal_pointer(&ret);
}
gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_client_dbus_call_sync_void(NMClient *self,
GCancellable *cancellable,
const char *object_path,
const char *interface_name,
const char *method_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GDBusCallFlags flags,
int timeout_msec,
gboolean strip_dbus_error,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GError **error)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
gs_unref_variant GVariant *ret = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = _nm_client_dbus_call_sync(self,
cancellable,
object_path,
interface_name,
method_name,
parameters,
G_VARIANT_TYPE("()"),
flags,
timeout_msec,
strip_dbus_error,
error);
return !!ret;
}
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_client_set_property_sync_legacy(NMClient *self,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const char *object_path,
const char *interface_name,
const char *property_name,
const char *format_string,
...)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
GVariant *val;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_variant GVariant *ret = NULL;
va_list ap;
nm_assert(!self || NM_IS_CLIENT(self));
nm_assert(interface_name);
nm_assert(property_name);
nm_assert(format_string);
if (!self)
return;
priv = NM_CLIENT_GET_PRIVATE(self);
if (!priv->name_owner)
return;
va_start(ap, format_string);
val = g_variant_new_va(format_string, NULL, &ap);
va_end(ap);
nm_assert(val);
/* A synchronous D-Bus call that is not cancellable an ignores the return value.
* This function only exists for backward compatibility. */
ret = g_dbus_connection_call_sync(priv->dbus_connection,
priv->name_owner,
object_path,
DBUS_INTERFACE_PROPERTIES,
"Set",
g_variant_new("(ssv)", interface_name, property_name, val),
NULL,
G_DBUS_CALL_FLAGS_NONE,
2000,
NULL,
NULL);
}
/*****************************************************************************/
#define _assert_main_context_is_current_source(self, x_context) \
G_STMT_START \
{ \
if (NM_MORE_ASSERTS > 0) { \
GSource *_source = g_main_current_source(); \
\
if (_source) { \
NMClientPrivate *_priv = NM_CLIENT_GET_PRIVATE(self); \
\
nm_assert(g_source_get_context(_source) == _priv->x_context); \
libnm: fix assertion in NMClient checking for current main context NMClient is strongly tied to the GMainContext with which it was created. Several operations must only be called from within the context. There was an assertion for that. However, creating (and init_async()) should be allowed to call not from within the GMainContext. So if the current context has no owner (is not acquired), then it's also OK. Fix the assertion for that. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2022-03-11 11:45:32 +01:00
nm_assert(nm_g_main_context_can_acquire(_priv->x_context)); \
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
} \
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
} \
} \
G_STMT_END
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
#define _assert_main_context_is_current_thread_default(self, x_context) \
G_STMT_START \
{ \
if (NM_MORE_ASSERTS > 0) { \
NMClientPrivate *_priv = NM_CLIENT_GET_PRIVATE(self); \
\
nm_assert((g_main_context_get_thread_default() ?: g_main_context_default()) \
== _priv->x_context); \
libnm: fix assertion in NMClient checking for current main context NMClient is strongly tied to the GMainContext with which it was created. Several operations must only be called from within the context. There was an assertion for that. However, creating (and init_async()) should be allowed to call not from within the GMainContext. So if the current context has no owner (is not acquired), then it's also OK. Fix the assertion for that. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2022-03-11 11:45:32 +01:00
nm_assert(nm_g_main_context_can_acquire(_priv->x_context)); \
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
} \
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
} \
G_STMT_END
/*****************************************************************************/
void
_nm_client_queue_notify_object(NMClient *self, gpointer nmobj, const GParamSpec *pspec)
{
NMObjectBase *base;
nm_assert(NM_IS_CLIENT(self));
nm_assert(NM_IS_OBJECT(nmobj) || NM_IS_CLIENT(nmobj));
base = (NMObjectBase *) nmobj;
libnm: don't emit property changed "notify" signal while destructing NMClient It seems to trip up gnome-control-center (rh #1778668). Just don't emit anymore signals once NMClient goes down.
2019-12-03 14:50:12 +01:00
if (base->is_disposing) {
/* Don't emit property changed signals once the NMClient
* instance is about to shut down. */
nm_assert(nmobj == self);
return;
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (c_list_is_empty(&base->queue_notify_lst)) {
c_list_link_tail(&NM_CLIENT_GET_PRIVATE(self)->queue_notify_lst_head,
&base->queue_notify_lst);
g_object_ref(nmobj);
g_object_freeze_notify(nmobj);
}
if (pspec)
g_object_notify_by_pspec(nmobj, (GParamSpec *) pspec);
}
/*****************************************************************************/
gpointer
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_client_notify_event_queue(NMClient *self,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
int priority,
NMClientNotifyEventCb callback,
gsize event_size)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientNotifyEvent *notify_event;
nm_assert(callback);
nm_assert(event_size > sizeof(NMClientNotifyEvent));
notify_event = g_malloc(event_size);
notify_event->priority = priority;
notify_event->callback = callback;
c_list_link_tail(&priv->notify_event_lst_head, &notify_event->lst);
priv->notify_event_lst_changed = TRUE;
return notify_event;
}
NMClientNotifyEventWithPtr *
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_client_notify_event_queue_with_ptr(NMClient *self,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
int priority,
NMClientNotifyEventWithPtrCb callback,
gpointer user_data)
{
NMClientNotifyEventWithPtr *notify_event;
notify_event = _nm_client_notify_event_queue(self,
priority,
(NMClientNotifyEventCb) callback,
sizeof(NMClientNotifyEventWithPtr));
notify_event->user_data = user_data;
return notify_event;
}
/*****************************************************************************/
typedef struct {
NMClientNotifyEvent parent;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GObject *source;
NMObject *obj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
guint signal_id;
} NMClientNotifyEventObjAddedRemove;
static void
_nm_client_notify_event_queue_emit_obj_signal_cb(NMClient *self, gpointer notify_event_base)
{
NMClientNotifyEventObjAddedRemove *notify_event = notify_event_base;
NML_NMCLIENT_LOG_T(
self,
"[%s] emit \"%s\" signal for %s",
NM_IS_CLIENT(notify_event->source) ? "nmclient" : _nm_object_get_path(notify_event->source),
g_signal_name(notify_event->signal_id),
_nm_object_get_path(notify_event->obj));
nm_assert(NM_IS_OBJECT(notify_event->source) || NM_IS_CLIENT(notify_event->source));
g_signal_emit(notify_event->source, notify_event->signal_id, 0, notify_event->obj);
g_object_unref(notify_event->obj);
g_object_unref(notify_event->source);
}
void
_nm_client_notify_event_queue_emit_obj_signal(NMClient *self,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GObject *source,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMObject *nmobj,
gboolean is_added /* or else removed */,
int prio_offset,
guint signal_id)
{
NMClientNotifyEventObjAddedRemove *notify_event;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(prio_offset >= 0);
nm_assert(prio_offset < 20);
nm_assert(NM_IS_OBJECT(source) || NM_IS_CLIENT(source));
nm_assert(NM_IS_OBJECT(nmobj));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (((NMObjectBase *) source)->is_disposing) {
nm_assert(NM_IS_CLIENT(source));
return;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
notify_event = _nm_client_notify_event_queue(
self,
is_added ? NM_CLIENT_NOTIFY_EVENT_PRIO_AFTER - 20 + prio_offset
: NM_CLIENT_NOTIFY_EVENT_PRIO_BEFORE + 20 - prio_offset,
_nm_client_notify_event_queue_emit_obj_signal_cb,
sizeof(NMClientNotifyEventObjAddedRemove));
notify_event->source = g_object_ref(source);
notify_event->obj = g_object_ref(nmobj);
notify_event->signal_id = signal_id;
}
/*****************************************************************************/
static int
_nm_client_notify_event_cmp(const CList *a, const CList *b, const void *user_data)
{
NM_CMP_DIRECT(c_list_entry(a, NMClientNotifyEvent, lst)->priority,
c_list_entry(b, NMClientNotifyEvent, lst)->priority);
return 0;
}
static void
_nm_client_notify_event_emit_parts(NMClient *self, int max_priority /* included! */)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientNotifyEvent *notify_event;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
while (TRUE) {
if (priv->notify_event_lst_changed) {
priv->notify_event_lst_changed = FALSE;
c_list_sort(&priv->notify_event_lst_head, _nm_client_notify_event_cmp, NULL);
}
notify_event = c_list_first_entry(&priv->notify_event_lst_head, NMClientNotifyEvent, lst);
if (!notify_event)
return;
if (notify_event->priority > max_priority)
return;
c_list_unlink_stale(&notify_event->lst);
notify_event->callback(self, notify_event);
g_free(notify_event);
}
}
static void
_nm_client_notify_event_emit(NMClient *self)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMObjectBase *base;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_notify_event_emit_parts(self, NM_CLIENT_NOTIFY_EVENT_PRIO_GPROP);
while (
(base = c_list_first_entry(&priv->queue_notify_lst_head, NMObjectBase, queue_notify_lst))) {
c_list_unlink(&base->queue_notify_lst);
g_object_thaw_notify(G_OBJECT(base));
g_object_unref(base);
}
_nm_client_notify_event_emit_parts(self, G_MAXINT);
}
/*****************************************************************************/
GDBusConnection *
_nm_client_get_dbus_connection(NMClient *self)
{
return NM_CLIENT_GET_PRIVATE(self)->dbus_connection;
}
const char *
_nm_client_get_dbus_name_owner(NMClient *self)
{
return NM_CLIENT_GET_PRIVATE(self)->name_owner;
}
GMainContext *
_nm_client_get_context_main(NMClient *self)
{
return NM_CLIENT_GET_PRIVATE(self)->main_context;
}
GMainContext *
_nm_client_get_context_dbus(NMClient *self)
{
return NM_CLIENT_GET_PRIVATE(self)->dbus_context;
}
libnm: add nm_client_get_main_context() function The NMClient is associated with a certain context. Add a getter function to give the context. The context is really not internal API of NMClient, that is because the user must iterate this context and be aware of it.
2019-11-26 13:31:11 +01:00
/**
* nm_client_get_main_context:
* @self: the #NMClient instance
*
* The #NMClient instance is permanently associated with the current
* thread default #GMainContext, referenced the time when the instance
* was created. To receive events, the user must iterate this context
* and can use it to synchronize access to the client.
*
* Note that even after #NMClient instance got destroyed, there might
* still be pending sources registered in the context. That means, to fully
* clean up, the user must continue iterating the context as long as
* the nm_client_get_context_busy_watcher() object is alive.
*
* Returns: (transfer none): the #GMainContext of the client.
*
* Since: 1.22
*/
GMainContext *
nm_client_get_main_context(NMClient *self)
{
g_return_val_if_fail(NM_IS_CLIENT(self), NULL);
return _nm_client_get_context_main(self);
}
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
/**
* nm_client_get_context_busy_watcher:
* @self: the NMClient instance.
*
* Returns: (transfer none): a GObject that stays alive as long as there are pending
* D-Bus operations.
*
* NMClient will schedule asynchronous D-Bus requests which will complete on
* the GMainContext associated with the instance. When destroying the NMClient
* instance, those requests are cancelled right away, however their pending requests are
* still outstanding and queued in the GMainContext. These outstanding callbacks
* keep the GMainContext alive. In order to fully release all resources,
* the user must keep iterating the main context until all these callbacks
* are handled. Of course, at this point no more actual callbacks will be invoked
libnm/doc: clarify GMainContext handling in NMClient documentation https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/819
2021-04-18 08:34:27 +02:00
* for the user, those are all cancelled internally.
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
*
* This just leaves one problem: how long does the user need to keep the
* GMainContext running to ensure everything is cleaned up? The answer is
* this GObject. Subscribe a weak reference to the returned object and keep
* iterating the main context until the object got unreferenced.
*
libnm/doc: clarify GMainContext handling in NMClient documentation https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/819
2021-04-18 08:34:27 +02:00
* Note that after the NMClient instance gets destroyed, all outstanding operations
* will be cancelled right away. That means, the user needs to iterate the #GMainContext
* a bit longer, but it is guaranteed that the cleanup happens soon after.
*
* The way of using the context-busy-watch, is by registering a weak pointer to
* see when it gets destroyed. That means, user code should not take additional
* references on this object to not keep it alive longer.
*
* If you plan to exit the program after releasing the NMClient instance
* you may not need to worry about these "leaks". Also, if you anyway plan to continue
* iterating the #GMainContext afterwards, then you don't need to care when exactly
* NMClient is gone completely.
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
*
* Since: 1.22
*/
GObject *
nm_client_get_context_busy_watcher(NMClient *self)
{
GObject *w;
g_return_val_if_fail(NM_IS_CLIENT(self), NULL);
w = NM_CLIENT_GET_PRIVATE(self)->context_busy_watcher;
libnm: expose nm_context_busy_watcher_integrate_source() as internal API for reuse This will also be useful for NMSecretAgentOld. The mechanics how NMClient handles the GMainContext and the context-busy-watcher apply really to every GObject that uses GDBusConnection and registers to signals. At least, as long as the API provides no shutdown/stop method, because that means shutdown/stop happens when unreferencing the instance, at which point pending operations get cancelled (but they cannot complete right away due to the nature of GTask and g_dbus_connection_call()). If there is a shutdown/stop API, then all pending operations could keep the instance alive, and the instance would sticks around (and keeps the GMainContext busy) until shutdown is completed. Basically, then the instance could be the context-busy-watcher itself. But in existing API which does not require the user to explicitly shutdown, that is not a feasible (backward compatible) addition. But the context-busy-watcher object is.
2019-12-17 13:59:15 +01:00
return g_object_get_qdata(w, nm_context_busy_watcher_quark()) ?: w;
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
struct udev *
_nm_client_get_udev(NMClient *self)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
if (G_UNLIKELY(!priv->udev_inited)) {
priv->udev_inited = TRUE;
/* for testing, we don't want to use udev in libnm. */
if (!nm_streq0(g_getenv("LIBNM_USE_NO_UDEV"), "1"))
priv->udev = udev_new();
}
return priv->udev;
}
/*****************************************************************************/
static void
_ASSERT_dbobj(NMLDBusObject *dbobj, NMClient *self)
{
#if NM_MORE_ASSERTS > 5
nm_assert(NM_IS_CLIENT(self));
nm_assert(NML_IS_DBUS_OBJECT(dbobj));
nm_assert(dbobj == g_hash_table_lookup(NM_CLIENT_GET_PRIVATE(self)->dbus_objects, dbobj));
#endif
}
static NMLDBusObject *
nml_dbus_object_new(NMRefString *dbus_path_take)
{
NMLDBusObject *dbobj;
nm_assert(NM_IS_REF_STRING(dbus_path_take));
dbobj = g_slice_new(NMLDBusObject);
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
*dbobj = (NMLDBusObject) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
.dbus_path = g_steal_pointer(&dbus_path_take),
.ref_count = 1,
.dbus_objects_lst = C_LIST_INIT(dbobj->dbus_objects_lst),
.iface_lst_head = C_LIST_INIT(dbobj->iface_lst_head),
.watcher_lst_head = C_LIST_INIT(dbobj->watcher_lst_head),
.obj_changed_lst = C_LIST_INIT(dbobj->obj_changed_lst),
.obj_state = NML_DBUS_OBJ_STATE_UNLINKED,
};
return dbobj;
}
NMLDBusObject *
nml_dbus_object_ref(NMLDBusObject *dbobj)
{
nm_assert(dbobj);
nm_assert(dbobj->ref_count > 0);
dbobj->ref_count++;
return dbobj;
}
void
nml_dbus_object_unref(NMLDBusObject *dbobj)
{
nm_assert(dbobj);
nm_assert(dbobj->ref_count > 0);
if (--dbobj->ref_count > 0)
return;
nm_assert(c_list_is_empty(&dbobj->obj_changed_lst));
nm_assert(c_list_is_empty(&dbobj->iface_lst_head));
nm_assert(c_list_is_empty(&dbobj->watcher_lst_head));
nm_assert(!dbobj->nmobj);
nm_ref_string_unref(dbobj->dbus_path);
nm_g_slice_free(dbobj);
}
static NMLDBusObjIfaceData *
nml_dbus_object_iface_data_get(NMLDBusObject *dbobj,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *dbus_iface_name,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean allow_create)
{
const NMLDBusMetaIface *meta_iface;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObjIfaceData *db_iface_data;
NMLDBusObjPropData *db_prop_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
guint count = 0;
guint i;
nm_assert(NML_IS_DBUS_OBJECT(dbobj));
nm_assert(dbus_iface_name);
#if NM_MORE_ASSERTS > 10
{
gboolean expect_well_known = TRUE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/* all well-known interfaces must come first in the list. */
c_list_for_each_entry (db_iface_data, &dbobj->iface_lst_head, iface_lst) {
if (db_iface_data->dbus_iface_is_wellknown == expect_well_known)
continue;
nm_assert(expect_well_known);
expect_well_known = FALSE;
}
}
#endif
meta_iface = nml_dbus_meta_iface_get(dbus_iface_name);
if (meta_iface) {
c_list_for_each_entry (db_iface_data, &dbobj->iface_lst_head, iface_lst) {
if (!db_iface_data->dbus_iface_is_wellknown)
break;
if (db_iface_data->iface_removed)
continue;
if (db_iface_data->dbus_iface.meta == meta_iface)
return db_iface_data;
count++;
}
} else {
shared: add c-list macros to iterate backwards
2020-11-04 17:26:50 +01:00
c_list_for_each_entry_prev (db_iface_data, &dbobj->iface_lst_head, iface_lst) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (db_iface_data->dbus_iface_is_wellknown)
break;
if (db_iface_data->iface_removed)
continue;
if (nm_streq(db_iface_data->dbus_iface.name->str, dbus_iface_name))
return db_iface_data;
count++;
}
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!allow_create)
return NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (count > 20) {
/* We track the list of interfaces that an object has in a linked list.
* That is efficient and convenient, if we assume that each object only has a small
* number of interfaces (which very much should be the case). Here, something is very
* odd, maybe there is a bug or the server side is misbehaving. Anyway, error out. */
return NULL;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
db_iface_data = g_malloc(
G_STRUCT_OFFSET(NMLDBusObjIfaceData, prop_datas)
+ (meta_iface ? (sizeof(NMLDBusObjPropData) * meta_iface->n_dbus_properties) : 0u));
if (meta_iface) {
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
*db_iface_data = (NMLDBusObjIfaceData) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
.dbus_iface.meta = meta_iface,
.dbus_iface_is_wellknown = TRUE,
.changed_prop_lst_head = C_LIST_INIT(db_iface_data->changed_prop_lst_head),
.iface_removed = FALSE,
};
db_prop_data = &db_iface_data->prop_datas[0];
for (i = 0; i < meta_iface->n_dbus_properties; i++, db_prop_data++) {
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
*db_prop_data = (NMLDBusObjPropData) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
.prop_data_value = NULL,
.changed_prop_lst = C_LIST_INIT(db_prop_data->changed_prop_lst),
};
}
c_list_link_front(&dbobj->iface_lst_head, &db_iface_data->iface_lst);
} else {
/* Intentionally don't initialize the other fields. We are not supposed
* to touch them, and a valgrind warning would be preferable. */
db_iface_data->dbus_iface.name = nm_ref_string_new(dbus_iface_name);
db_iface_data->dbus_iface_is_wellknown = FALSE;
db_iface_data->iface_removed = FALSE;
c_list_link_tail(&dbobj->iface_lst_head, &db_iface_data->iface_lst);
}
return db_iface_data;
}
static void
nml_dbus_obj_iface_data_destroy(NMLDBusObjIfaceData *db_iface_data)
{
guint i;
nm_assert(db_iface_data);
nm_assert(c_list_is_empty(&db_iface_data->iface_lst));
if (db_iface_data->dbus_iface_is_wellknown) {
for (i = 0; i < db_iface_data->dbus_iface.meta->n_dbus_properties; i++)
nm_g_variant_unref(db_iface_data->prop_datas[i].prop_data_value);
} else
nm_ref_string_unref(db_iface_data->dbus_iface.name);
g_free(db_iface_data);
}
gpointer
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nml_dbus_object_get_property_location(NMLDBusObject *dbobj,
const NMLDBusMetaIface *meta_iface,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaProperty *meta_property)
{
char *target_c;
all: fix various "-Wcast-align=strict" warnings The warning "-Wcast-align=strict" seems useful and will be enabled next. Fix places that currently cause the warning by using the new macro NM_CAST_ALIGN(). This macro also nm_assert()s that the alignment is correct.
2022-12-01 14:07:22 +01:00
target_c = ((gpointer) dbobj->nmobj);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (meta_iface->base_struct_offset > 0)
all: fix various "-Wcast-align=strict" warnings The warning "-Wcast-align=strict" seems useful and will be enabled next. Fix places that currently cause the warning by using the new macro NM_CAST_ALIGN(). This macro also nm_assert()s that the alignment is correct.
2022-12-01 14:07:22 +01:00
target_c = *NM_CAST_ALIGN(gpointer, &target_c[meta_iface->base_struct_offset]);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return &target_c[meta_property->prop_struct_offset];
}
static void
nml_dbus_object_set_obj_state(NMLDBusObject *dbobj, NMLDBusObjState obj_state, NMClient *self)
{
NMClientPrivate *priv;
nm_assert(NM_IS_CLIENT(self));
nm_assert(NML_IS_DBUS_OBJECT(dbobj));
#if NM_MORE_ASSERTS > 10
priv = NM_CLIENT_GET_PRIVATE(self);
switch (dbobj->obj_state) {
case NML_DBUS_OBJ_STATE_UNLINKED:
nm_assert(c_list_is_empty(&dbobj->dbus_objects_lst));
break;
case NML_DBUS_OBJ_STATE_WATCHED_ONLY:
nm_assert(
c_list_contains(&priv->dbus_objects_lst_head_watched_only, &dbobj->dbus_objects_lst));
break;
case NML_DBUS_OBJ_STATE_ON_DBUS:
nm_assert(c_list_contains(&priv->dbus_objects_lst_head_on_dbus, &dbobj->dbus_objects_lst));
break;
case NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY:
nm_assert(c_list_contains(&priv->dbus_objects_lst_head_with_nmobj_not_ready,
&dbobj->dbus_objects_lst));
break;
case NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY:
nm_assert(c_list_contains(&priv->dbus_objects_lst_head_with_nmobj_ready,
&dbobj->dbus_objects_lst));
break;
}
#endif
if (dbobj->obj_state == obj_state)
return;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(self,
"[%s]: set D-Bus object state %s",
dbobj->dbus_path->str,
nml_dbus_obj_state_to_string(obj_state));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv = NM_CLIENT_GET_PRIVATE(self);
dbobj->obj_state = obj_state;
switch (obj_state) {
case NML_DBUS_OBJ_STATE_UNLINKED:
c_list_unlink(&dbobj->dbus_objects_lst);
c_list_unlink(&dbobj->obj_changed_lst);
dbobj->obj_changed_type = NML_DBUS_OBJ_CHANGED_TYPE_NONE;
break;
case NML_DBUS_OBJ_STATE_WATCHED_ONLY:
nm_c_list_move_tail(&priv->dbus_objects_lst_head_watched_only, &dbobj->dbus_objects_lst);
break;
case NML_DBUS_OBJ_STATE_ON_DBUS:
nm_c_list_move_tail(&priv->dbus_objects_lst_head_on_dbus, &dbobj->dbus_objects_lst);
break;
case NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY:
nm_c_list_move_tail(&priv->dbus_objects_lst_head_with_nmobj_not_ready,
&dbobj->dbus_objects_lst);
break;
case NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY:
nm_c_list_move_tail(&priv->dbus_objects_lst_head_with_nmobj_ready,
&dbobj->dbus_objects_lst);
break;
default:
nm_assert_not_reached();
}
}
/*****************************************************************************/
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nml_dbus_object_obj_changed_link(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObjChangedType changed_type)
{
nm_assert(NM_IS_CLIENT(self));
nm_assert(NML_IS_DBUS_OBJECT(dbobj));
nm_assert(changed_type != NML_DBUS_OBJ_CHANGED_TYPE_NONE);
libnm: add code comments to NMClient
2022-01-27 22:09:26 +01:00
/* Links @dbobj in the "obj_changed_lst", with the new "changed_type". */
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!NM_FLAGS_ALL((NMLDBusObjChangedType) dbobj->obj_changed_type, changed_type))
libnm: minor cleanup of libnm trace logging
2020-01-30 16:54:28 +01:00
NML_NMCLIENT_LOG_T(self,
"[%s]: changed-type 0x%02x linked",
dbobj->dbus_path->str,
(guint) changed_type);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (dbobj->obj_changed_type == NML_DBUS_OBJ_CHANGED_TYPE_NONE) {
NMClientPrivate *priv;
/* We set the changed-type flag. Need to queue the object in the
* changed list. */
nm_assert(c_list_is_empty(&dbobj->obj_changed_lst));
priv = NM_CLIENT_GET_PRIVATE(self);
c_list_link_tail(&priv->obj_changed_lst_head, &dbobj->obj_changed_lst);
} else {
/* The object has changes flags and must be linked already. Note that
* this may be priv->obj_changed_lst_head, or a temporary list on the
* stack.
*
* This dance with the temporary list is done to ensure we can enqueue
* objects while we process the changes. */
nm_assert(!c_list_is_empty(&dbobj->obj_changed_lst));
}
dbobj->obj_changed_type |= changed_type;
nm_assert(NM_FLAGS_ALL(dbobj->obj_changed_type, changed_type));
}
static NMLDBusObjChangedType
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nml_dbus_object_obj_changed_consume(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObjChangedType changed_type)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObjChangedType changed_type_res;
libnm: add code comments to NMClient
2022-01-27 22:09:26 +01:00
/* We have @dbobj which has some "obj_changed_type" set (consequently,
* it's linked in the "obj_changed_lst"). Here we consume the @changed_type,
* meaning, to clear those flags from "obj_change_type" (and return
* the flags that were cleared/present or NONE, if the current object
* doesn't have these changed-types. */
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
nm_assert(NML_IS_DBUS_OBJECT(dbobj));
nm_assert(changed_type != NML_DBUS_OBJ_CHANGED_TYPE_NONE);
nm_assert(dbobj->obj_changed_type != NML_DBUS_OBJ_CHANGED_TYPE_NONE);
nm_assert(!c_list_is_empty(&dbobj->obj_changed_lst));
changed_type_res = dbobj->obj_changed_type & changed_type;
dbobj->obj_changed_type &= ~changed_type;
if (dbobj->obj_changed_type == NML_DBUS_OBJ_CHANGED_TYPE_NONE) {
libnm: add code comments to NMClient
2022-01-27 22:09:26 +01:00
/* No other "obj_change_type" left. Unlink the object from the
* "changed_type_list". */
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_unlink(&dbobj->obj_changed_lst);
nm_assert(changed_type_res != NML_DBUS_OBJ_CHANGED_TYPE_NONE);
libnm: minor cleanup of libnm trace logging
2020-01-30 16:54:28 +01:00
NML_NMCLIENT_LOG_T(self,
"[%s]: changed-type 0x%02x consumed",
dbobj->dbus_path->str,
(guint) changed_type_res);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return changed_type_res;
}
priv = NM_CLIENT_GET_PRIVATE(self);
libnm: add code comments to NMClient
2022-01-27 22:09:26 +01:00
/* Actually, at this point, @dbobj is not linked in priv->obj_changed_lst_head,
* instead, it's linked on a temporary list. As we still have changes left after
* consuming "changed_type", we move it to priv->obj_changed_lst_head. */
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(!c_list_contains(&priv->obj_changed_lst_head, &dbobj->obj_changed_lst));
nm_c_list_move_tail(&priv->obj_changed_lst_head, &dbobj->obj_changed_lst);
libnm: minor cleanup of libnm trace logging
2020-01-30 16:54:28 +01:00
NML_NMCLIENT_LOG_T(self,
"[%s]: changed-type 0x%02x consumed (still has 0x%02x)",
dbobj->dbus_path->str,
(guint) changed_type_res,
(guint) dbobj->obj_changed_type);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return changed_type_res;
}
static gboolean
nml_dbus_object_obj_changed_any_linked(NMClient *self, NMLDBusObjChangedType changed_type)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(changed_type != NML_DBUS_OBJ_CHANGED_TYPE_NONE);
c_list_for_each_entry (dbobj, &priv->obj_changed_lst_head, obj_changed_lst) {
nm_assert(dbobj->obj_changed_type != NML_DBUS_OBJ_CHANGED_TYPE_NONE);
if (NM_FLAGS_ANY(dbobj->obj_changed_type, changed_type))
return TRUE;
}
return FALSE;
}
/*****************************************************************************/
static void
_dbobjs_notify_watchers_for_dbobj(NMClient *self, NMLDBusObject *dbobj)
{
NMLDBusObjWatcher *obj_watcher;
NMLDBusObjWatcher *obj_watcher_safe;
c_list_for_each_entry_safe (obj_watcher,
obj_watcher_safe,
&dbobj->watcher_lst_head,
_priv.watcher_lst)
obj_watcher->_priv.notify_fcn(self, obj_watcher);
}
static gboolean
_dbobjs_check_dbobj_ready(NMClient *self, NMLDBusObject *dbobj)
{
nm_assert(NM_IS_CLIENT(self));
nm_assert(NML_IS_DBUS_OBJECT(dbobj));
nm_assert(G_IS_OBJECT(dbobj->nmobj));
nm_assert(NM_IS_OBJECT(dbobj->nmobj) || NM_IS_CLIENT(dbobj->nmobj));
nm_assert(NM_IN_SET((NMLDBusObjState) dbobj->obj_state,
NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY,
NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY));
if (G_LIKELY(dbobj->obj_state == NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY))
return TRUE;
if (!NM_OBJECT_GET_CLASS(dbobj->nmobj)->is_ready(NM_OBJECT(dbobj->nmobj)))
return FALSE;
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY, self);
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
_dbobjs_notify_watchers_for_dbobj(self, dbobj);
return TRUE;
}
void
_nm_client_notify_object_changed(NMClient *self, NMLDBusObject *dbobj)
{
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
_dbobjs_notify_watchers_for_dbobj(self, dbobj);
}
/*****************************************************************************/
static NMLDBusObject *
_dbobjs_dbobj_get_r(NMClient *self, NMRefString *dbus_path_r)
{
nm_assert(NM_IS_REF_STRING(dbus_path_r));
return g_hash_table_lookup(NM_CLIENT_GET_PRIVATE(self)->dbus_objects, &dbus_path_r);
}
static NMLDBusObject *
_dbobjs_dbobj_get_s(NMClient *self, const char *dbus_path)
{
nm_auto_ref_string NMRefString *dbus_path_r = NULL;
nm_assert(dbus_path);
dbus_path_r = nm_ref_string_new(dbus_path);
return _dbobjs_dbobj_get_r(self, dbus_path_r);
}
static NMLDBusObject *
_dbobjs_dbobj_create(NMClient *self, NMRefString *dbus_path_take)
{
nm_auto_ref_string NMRefString *dbus_path = g_steal_pointer(&dbus_path_take);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(!_dbobjs_dbobj_get_r(self, dbus_path));
dbobj = nml_dbus_object_new(g_steal_pointer(&dbus_path));
if (!g_hash_table_add(priv->dbus_objects, dbobj))
nm_assert_not_reached();
return dbobj;
}
static NMLDBusObject *
_dbobjs_dbobj_get_or_create(NMClient *self, NMRefString *dbus_path_take)
{
nm_auto_ref_string NMRefString *dbus_path = g_steal_pointer(&dbus_path_take);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbobj = _dbobjs_dbobj_get_r(self, dbus_path);
if (dbobj)
return dbobj;
return _dbobjs_dbobj_create(self, g_steal_pointer(&dbus_path));
}
static NMLDBusObject *
_dbobjs_get_nmobj(NMClient *self, const char *dbus_path, GType gtype)
{
NMLDBusObject *dbobj;
nm_assert(gtype == G_TYPE_NONE || g_type_is_a(gtype, NM_TYPE_OBJECT));
dbobj = _dbobjs_dbobj_get_s(self, dbus_path);
if (!dbobj)
return NULL;
if (!dbobj->nmobj)
return NULL;
if (gtype != G_TYPE_NONE && !g_type_is_a(G_OBJECT_TYPE(dbobj->nmobj), gtype))
return NULL;
return dbobj;
}
static gpointer
_dbobjs_get_nmobj_unpack_visible(NMClient *self, const char *dbus_path, GType gtype)
{
NMLDBusObject *dbobj;
dbobj = _dbobjs_get_nmobj(self, dbus_path, gtype);
if (!dbobj)
return NULL;
if (dbobj->obj_state != NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY)
return NULL;
return dbobj->nmobj;
}
/*****************************************************************************/
static gpointer
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_dbobjs_obj_watcher_register_o(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObjWatchNotifyFcn notify_fcn,
gsize struct_size)
{
NMLDBusObjWatcher *obj_watcher;
nm_assert(NM_IS_CLIENT(self));
_ASSERT_dbobj(dbobj, self);
nm_assert(notify_fcn);
nm_assert(struct_size > sizeof(NMLDBusObjWatcher));
obj_watcher = g_malloc(struct_size);
obj_watcher->dbobj = dbobj;
obj_watcher->_priv.notify_fcn = notify_fcn;
/* we must enqueue the item in the front of the list. That is, because while
* invoking notify_fcn(), we iterate the watchers front-to-end. As we want to
* allow the callee to register new watches and unregister itself, this is
* the right way to do it. */
c_list_link_front(&dbobj->watcher_lst_head, &obj_watcher->_priv.watcher_lst);
return obj_watcher;
}
static gpointer
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_dbobjs_obj_watcher_register_r(NMClient *self,
NMRefString *dbus_path_take,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObjWatchNotifyFcn notify_fcn,
gsize struct_size)
{
nm_auto_ref_string NMRefString *dbus_path = g_steal_pointer(&dbus_path_take);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
nm_assert(notify_fcn);
dbobj = _dbobjs_dbobj_get_or_create(self, g_steal_pointer(&dbus_path));
if (dbobj->obj_state == NML_DBUS_OBJ_STATE_UNLINKED)
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_WATCHED_ONLY, self);
return _dbobjs_obj_watcher_register_o(self, dbobj, notify_fcn, struct_size);
}
static void
_dbobjs_obj_watcher_unregister(NMClient *self, gpointer obj_watcher_base)
{
NMLDBusObjWatcher *obj_watcher = obj_watcher_base;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
nm_assert(obj_watcher);
nm_assert(NML_IS_DBUS_OBJECT(obj_watcher->dbobj));
nm_assert(g_hash_table_lookup(NM_CLIENT_GET_PRIVATE(self)->dbus_objects, obj_watcher->dbobj)
== obj_watcher->dbobj);
nm_assert(
c_list_contains(&obj_watcher->dbobj->watcher_lst_head, &obj_watcher->_priv.watcher_lst));
c_list_unlink(&obj_watcher->_priv.watcher_lst);
dbobj = obj_watcher->dbobj;
g_free(obj_watcher);
if (c_list_is_empty(&dbobj->iface_lst_head) && c_list_is_empty(&dbobj->watcher_lst_head)) {
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
NML_NMCLIENT_LOG_T(self, "[%s]: drop D-Bus watcher", dbobj->dbus_path->str);
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_UNLINKED, self);
if (!g_hash_table_steal(priv->dbus_objects, dbobj))
nm_assert_not_reached();
nml_dbus_object_unref(dbobj);
}
}
/*****************************************************************************/
typedef struct {
NMLDBusObjWatcher parent;
NMLDBusPropertyO *pr_o;
} PropertyOData;
gpointer
nml_dbus_property_o_get_obj(NMLDBusPropertyO *pr_o)
{
nm_assert(!pr_o->nmobj || nml_dbus_property_o_is_ready(pr_o));
return pr_o->nmobj;
}
gboolean
nml_dbus_property_o_is_ready(const NMLDBusPropertyO *pr_o)
{
return pr_o->is_ready || !pr_o->owner_dbobj;
}
libnm: hide NMActiveConnection until NMRemoteConnection is ready Generally, libnm's NMClient cache only wants to expose NMObjects that are fully initalized. Most objects don't require anything special, except NMRemoteConnection which waits for the GetSettings() call to complete. NMObjects reference each other. For example, NMActiveConnection references NMDevice and NMRemoteConnection. There is a desire that an object is only ready, if the objects that it references are ready too. In practice that is not done, because usually every objects references other objects, that means all objects would be declared as non-ready as long as any of them is still initializing. That does not seem desirable. Instead, most objects (except NMRemoteConnection and now NMActiveConnection) are considered ready and visible, once their first notification completes. In case the objects reference any object that is not yet ready, the references is NULL (but the source object is visible already). This is also done this way, to cope with cycles where objects reference each other. In practice, such cycles should not be exposed by NetworkManager. However, libnm should be robust against that. This has the undesired effect that when you call AddAndActivate(), then the NMActiveConnection might already be visible while its NMRemoteConnection isn't. That means, ac.get_connection() will initially return NULL, until the remote connection becomes ready. Also add a special handling that NMActiveConnection waits for their NMRemoteConnection to be ready, before being ready itself. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-31 00:54:46 +01:00
gboolean
nml_dbus_property_o_is_ready_fully(const NMLDBusPropertyO *pr_o)
{
return !pr_o->owner_dbobj || !pr_o->obj_watcher || pr_o->nmobj;
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static void
nml_dbus_property_o_notify_changed(NMLDBusPropertyO *pr_o, NMClient *self)
{
const NMLDBusPropertVTableO *vtable;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GObject *nmobj = NULL;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean is_ready = TRUE;
gboolean changed_ready;
GType gtype;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(pr_o);
nm_assert(NM_IS_CLIENT(self));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_o->owner_dbobj)
return;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_o->is_changed) {
if (pr_o->is_ready)
return;
goto done;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_o->is_changed = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_o->obj_watcher)
goto done;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_o->obj_watcher->dbobj->nmobj) {
if (pr_o->obj_watcher->dbobj->obj_state >= NML_DBUS_OBJ_STATE_ON_DBUS) {
NML_NMCLIENT_LOG_W(
self,
"[%s]: property %s references %s but object is not created",
pr_o->owner_dbobj->dbus_path->str,
pr_o->meta_iface->dbus_properties[pr_o->dbus_property_idx].dbus_property_name,
pr_o->obj_watcher->dbobj->dbus_path->str);
} else {
NML_NMCLIENT_LOG_E(
self,
"[%s]: property %s references %s but object is not present on D-Bus",
pr_o->owner_dbobj->dbus_path->str,
pr_o->meta_iface->dbus_properties[pr_o->dbus_property_idx].dbus_property_name,
pr_o->obj_watcher->dbobj->dbus_path->str);
}
goto done;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
vtable = pr_o->meta_iface->dbus_properties[pr_o->dbus_property_idx].extra.property_vtable_o;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gtype = vtable->get_o_type_fcn();
if (!g_type_is_a(G_OBJECT_TYPE(pr_o->obj_watcher->dbobj->nmobj), gtype)) {
NML_NMCLIENT_LOG_E(
self,
"[%s]: property %s references %s with unexpected GObject type %s instead of %s",
pr_o->owner_dbobj->dbus_path->str,
pr_o->meta_iface->dbus_properties[pr_o->dbus_property_idx].dbus_property_name,
pr_o->obj_watcher->dbobj->dbus_path->str,
G_OBJECT_TYPE_NAME(pr_o->obj_watcher->dbobj->nmobj),
g_type_name(gtype));
goto done;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (pr_o->obj_watcher->dbobj == pr_o->owner_dbobj) {
NML_NMCLIENT_LOG_W(
self,
"[%s]: property %s references itself",
pr_o->owner_dbobj->dbus_path->str,
pr_o->meta_iface->dbus_properties[pr_o->dbus_property_idx].dbus_property_name);
nmobj = pr_o->owner_dbobj->nmobj;
goto done;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_o->block_is_changed = TRUE;
is_ready = _dbobjs_check_dbobj_ready(self, pr_o->obj_watcher->dbobj);
pr_o->block_is_changed = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!is_ready) {
is_ready = vtable->is_always_ready;
goto done;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nmobj = pr_o->obj_watcher->dbobj->nmobj;
done:
changed_ready = FALSE;
if (!pr_o->is_ready && is_ready) {
pr_o->is_ready = TRUE;
changed_ready = TRUE;
}
if (pr_o->nmobj != nmobj) {
pr_o->nmobj = nmobj;
_nm_client_queue_notify_object(
self,
pr_o->owner_dbobj->nmobj,
pr_o->meta_iface->obj_properties
[pr_o->meta_iface->dbus_properties[pr_o->dbus_property_idx].obj_properties_idx]);
}
if (changed_ready && pr_o->owner_dbobj->obj_state == NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY)
nml_dbus_object_obj_changed_link(self, pr_o->owner_dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
}
void
nml_dbus_property_o_notify_changed_many(NMLDBusPropertyO *ptr, guint len, NMClient *self)
{
while (len-- > 0)
nml_dbus_property_o_notify_changed(ptr++, self);
}
static void
nml_dbus_property_o_notify_watch_cb(NMClient *self, gpointer obj_watcher)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
PropertyOData *pr_o_data = obj_watcher;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusPropertyO *pr_o = pr_o_data->pr_o;
nm_assert(pr_o->obj_watcher == obj_watcher);
if (!pr_o->block_is_changed && !pr_o->is_changed) {
pr_o->is_changed = TRUE;
nml_dbus_object_obj_changed_link(self, pr_o->owner_dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
}
}
nm-client: expose the method nml_dbus_property_o_notify()
2023-02-01 10:13:14 -05:00
NMLDBusNotifyUpdatePropFlags
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nml_dbus_property_o_notify(NMClient *self,
NMLDBusPropertyO *pr_o,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface *meta_iface,
guint dbus_property_idx,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *value)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
const char *dbus_path = NULL;
gboolean changed = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_o->owner_dbobj) {
nm_assert(!pr_o->meta_iface);
nm_assert(pr_o->dbus_property_idx == 0);
nm_assert(!pr_o->is_ready);
pr_o->owner_dbobj = dbobj;
pr_o->meta_iface = meta_iface;
pr_o->dbus_property_idx = dbus_property_idx;
} else {
nm_assert(pr_o->owner_dbobj == dbobj);
nm_assert(pr_o->meta_iface == meta_iface);
nm_assert(pr_o->dbus_property_idx == dbus_property_idx);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (value)
dbus_path = nm_dbus_path_not_empty(g_variant_get_string(value, NULL));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (pr_o->obj_watcher
&& (!dbus_path || !nm_streq(dbus_path, pr_o->obj_watcher->dbobj->dbus_path->str))) {
_dbobjs_obj_watcher_unregister(self, g_steal_pointer(&pr_o->obj_watcher));
changed = TRUE;
}
if (!pr_o->obj_watcher && dbus_path) {
pr_o->obj_watcher = _dbobjs_obj_watcher_register_r(self,
nm_ref_string_new(dbus_path),
nml_dbus_property_o_notify_watch_cb,
sizeof(PropertyOData));
((PropertyOData *) pr_o->obj_watcher)->pr_o = pr_o;
changed = TRUE;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed && !pr_o->is_changed) {
pr_o->is_changed = TRUE;
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NONE;
}
void
nml_dbus_property_o_clear(NMLDBusPropertyO *pr_o, NMClient *self)
{
if (pr_o->obj_watcher) {
nm_assert(NM_IS_CLIENT(self));
_dbobjs_obj_watcher_unregister(self, g_steal_pointer(&pr_o->obj_watcher));
}
if (pr_o->nmobj && pr_o->owner_dbobj && pr_o->owner_dbobj->nmobj) {
_nm_client_queue_notify_object(
self,
pr_o->owner_dbobj->nmobj,
pr_o->meta_iface->obj_properties
[pr_o->meta_iface->dbus_properties[pr_o->dbus_property_idx].obj_properties_idx]);
}
pr_o->owner_dbobj = NULL;
pr_o->meta_iface = NULL;
pr_o->dbus_property_idx = 0;
libnm: fix assertion for cleaning up nml_dbus_property_o_notify() Usually, the nmobj never gets reused for one dbobj. That means, we really don't expect a nml_dbus_property_o_notify() for a property that was already cleared. However, that is for example not the case with NMClient itself. As NetworkManager gets restarted, the name owner gets lost, the property cleared but afterwards it might get notified again. That means, nml_dbus_property_o_notify() and nml_dbus_property_o_clear() must work well together, otherwise a sequence of nml_dbus_property_o_notify() nml_dbus_property_o_clear() nml_dbus_property_o_notify() leads to an assertion failure "nm_assert (!pr_o->is_ready)". Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2019-11-26 12:31:23 +01:00
pr_o->is_ready = FALSE;
libnm: fix dangling pointer in "o" properties when unregistering NMObject When NMClient gets destroyed, it unrefs all NMObject. We need to unbreak cycles then, and the property getters must return NULL. In particular, for "o" type properties (NMLDBusPropertyO), this was not done correctly. For example, calling nm_device_get_active_connection() while/after destroying the NMClient can give a dangling pointer and assertion failure. This will also be covered by test_activate_virtual(). Probably a similar issue can happen, when a D-Bus object gets removed (without destroying NMClient altogether). The fix is that nml_dbus_property_o_clear() needs to clear "nmobj". That is correct, because the pointer is no longer valid and should not be there. And the unit test shows that in fact a pointer is left there, and clearing it fixes it. That was different from an earlier attempt to fix this (in commit 62b2aa85e875 ('Revert "libnm: fix dangling pointer in public API while destructing NMClient"')), where clearing the pointer at a different place broke things. That attempt was wrong, because nml_dbus_property_o_notify_changed() needs to be the one that sets/clears nmobj field during a regular update. But the case here is not a regular update, nml_dbus_property_o_clear() happens during unregister/cleanup, and then we need to clear the pointer. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://bugzilla.redhat.com/show_bug.cgi?id=2039331 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/896 See-also: https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1064 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1075
2022-01-28 08:24:36 +01:00
pr_o->nmobj = NULL;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
void
nml_dbus_property_o_clear_many(NMLDBusPropertyO *pr_o, guint len, NMClient *self)
{
while (len-- > 0)
nml_dbus_property_o_clear(pr_o++, self);
}
/*****************************************************************************/
typedef struct _NMLDBusPropertyAOData {
NMLDBusObjWatcher obj_watcher;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusPropertyAO *parent;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
CList data_lst;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GObject *nmobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
struct _NMLDBusPropertyAOData *changed_next;
bool is_ready : 1;
bool is_notified : 1;
bool is_changed : 1;
bool block_is_changed : 1;
} PropertyAOData;
static void
_ASSERT_pr_ao(NMLDBusPropertyAO *pr_ao)
{
nm_assert(pr_ao);
#if NM_MORE_ASSERTS > 10
if (pr_ao->owner_dbobj) {
guint n_not_ready = 0;
guint n_is_changed = 0;
guint n_is_changed_2;
PropertyAOData *pr_ao_data;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_for_each_entry (pr_ao_data, &pr_ao->data_lst_head, data_lst) {
if (pr_ao_data->is_changed)
n_is_changed++;
if (!pr_ao_data->is_ready)
n_not_ready++;
}
nm_assert(n_not_ready == pr_ao->n_not_ready);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
n_is_changed_2 = 0;
pr_ao_data = pr_ao->changed_head;
while (pr_ao_data) {
nm_assert(pr_ao_data->is_changed);
n_is_changed_2++;
pr_ao_data = pr_ao_data->changed_next;
}
nm_assert(n_is_changed == n_is_changed_2);
}
#endif
}
static gboolean
nml_dbus_property_ao_notify_changed_ao(PropertyAOData *pr_ao_data,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean is_added /* or else removed */)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusPropertyAO *pr_ao;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusPropertVTableAO *vtable;
if (!pr_ao_data->nmobj)
return FALSE;
nm_assert(pr_ao_data->is_ready);
if (is_added) {
if (pr_ao_data->is_notified)
return FALSE;
pr_ao_data->is_notified = TRUE;
} else {
if (!pr_ao_data->is_notified)
return FALSE;
pr_ao_data->is_notified = FALSE;
}
pr_ao = pr_ao_data->parent;
vtable = pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].extra.property_vtable_ao;
if (vtable->notify_changed_ao)
vtable->notify_changed_ao(pr_ao, self, NM_OBJECT(pr_ao_data->nmobj), is_added);
return TRUE;
}
const GPtrArray *
nml_dbus_property_ao_get_objs_as_ptrarray(NMLDBusPropertyAO *pr_ao)
{
if (!pr_ao->arr) {
PropertyAOData *pr_ao_data;
gsize n;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
n = 0;
if (pr_ao->owner_dbobj) {
c_list_for_each_entry (pr_ao_data, &pr_ao->data_lst_head, data_lst) {
if (pr_ao_data->nmobj)
n++;
}
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_ao->arr = g_ptr_array_new_full(n, g_object_unref);
if (pr_ao->owner_dbobj) {
c_list_for_each_entry (pr_ao_data, &pr_ao->data_lst_head, data_lst) {
if (pr_ao_data->nmobj)
g_ptr_array_add(pr_ao->arr, g_object_ref(pr_ao_data->nmobj));
}
}
}
return pr_ao->arr;
}
gboolean
nml_dbus_property_ao_is_ready(const NMLDBusPropertyAO *pr_ao)
{
return pr_ao->n_not_ready == 0;
}
static void
nml_dbus_property_ao_notify_changed(NMLDBusPropertyAO *pr_ao, NMClient *self)
{
gboolean changed_prop = FALSE;
gboolean changed_ready = FALSE;
PropertyAOData *pr_ao_data;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
_ASSERT_pr_ao(pr_ao);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_ao->owner_dbobj)
return;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_ao->is_changed) {
if (pr_ao->n_not_ready == 0)
return;
goto done;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_ao->is_changed = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
while (pr_ao->changed_head) {
const NMLDBusPropertVTableAO *vtable;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GObject *nmobj = NULL;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean is_ready = TRUE;
GType gtype;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_ao_data = g_steal_pointer(&pr_ao->changed_head);
nm_assert(pr_ao_data->is_changed);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_ao->changed_head = pr_ao_data->changed_next;
pr_ao_data->is_changed = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_ao_data->obj_watcher.dbobj->nmobj) {
if (pr_ao_data->obj_watcher.dbobj->obj_state >= NML_DBUS_OBJ_STATE_ON_DBUS) {
NML_NMCLIENT_LOG_W(
self,
"[%s]: property %s references %s but object is not created",
pr_ao->owner_dbobj->dbus_path->str,
pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].dbus_property_name,
pr_ao_data->obj_watcher.dbobj->dbus_path->str);
} else {
NML_NMCLIENT_LOG_E(
self,
"[%s]: property %s references %s but object is not present on D-Bus",
pr_ao->owner_dbobj->dbus_path->str,
pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].dbus_property_name,
pr_ao_data->obj_watcher.dbobj->dbus_path->str);
}
goto done_pr_ao_data;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
vtable =
pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].extra.property_vtable_ao;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gtype = vtable->get_o_type_fcn();
if (!g_type_is_a(G_OBJECT_TYPE(pr_ao_data->obj_watcher.dbobj->nmobj), gtype)) {
NML_NMCLIENT_LOG_E(
self,
"[%s]: property %s references %s with unexpected GObject type %s instead of %s",
pr_ao->owner_dbobj->dbus_path->str,
pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].dbus_property_name,
pr_ao_data->obj_watcher.dbobj->dbus_path->str,
G_OBJECT_TYPE_NAME(pr_ao_data->obj_watcher.dbobj->nmobj),
g_type_name(gtype));
goto done_pr_ao_data;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (pr_ao_data->obj_watcher.dbobj == pr_ao->owner_dbobj) {
NML_NMCLIENT_LOG_W(
self,
"[%s]: property %s references itself",
pr_ao->owner_dbobj->dbus_path->str,
pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].dbus_property_name);
nmobj = pr_ao->owner_dbobj->nmobj;
goto done_pr_ao_data;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_ao_data->block_is_changed = TRUE;
is_ready = _dbobjs_check_dbobj_ready(self, pr_ao_data->obj_watcher.dbobj);
pr_ao_data->block_is_changed = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!is_ready) {
is_ready = vtable->is_always_ready;
goto done_pr_ao_data;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (vtable->check_nmobj_visible_fcn
&& !vtable->check_nmobj_visible_fcn(pr_ao_data->obj_watcher.dbobj->nmobj)) {
is_ready = TRUE;
goto done_pr_ao_data;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nmobj = pr_ao_data->obj_watcher.dbobj->nmobj;
done_pr_ao_data:
if (!pr_ao_data->is_ready && is_ready) {
nm_assert(pr_ao->n_not_ready > 0);
pr_ao->n_not_ready--;
pr_ao_data->is_ready = TRUE;
changed_ready = TRUE;
}
if (pr_ao_data->nmobj != nmobj) {
if (nml_dbus_property_ao_notify_changed_ao(pr_ao_data, self, FALSE))
changed_prop = TRUE;
pr_ao_data->nmobj = nmobj;
}
if (!pr_ao_data->is_notified) {
if (nml_dbus_property_ao_notify_changed_ao(pr_ao_data, self, TRUE))
changed_prop = TRUE;
}
}
_ASSERT_pr_ao(pr_ao);
done:
if (changed_prop) {
nm_clear_pointer(&pr_ao->arr, g_ptr_array_unref);
_nm_client_queue_notify_object(
self,
pr_ao->owner_dbobj->nmobj,
pr_ao->meta_iface->obj_properties
[pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].obj_properties_idx]);
}
if (changed_ready && pr_ao->n_not_ready == 0
&& pr_ao->owner_dbobj->obj_state == NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY)
nml_dbus_object_obj_changed_link(self, pr_ao->owner_dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
}
void
nml_dbus_property_ao_notify_changed_many(NMLDBusPropertyAO *ptr, guint len, NMClient *self)
{
while (len-- > 0)
nml_dbus_property_ao_notify_changed(ptr++, self);
}
static void
nml_dbus_property_ao_notify_watch_cb(NMClient *self, gpointer obj_watcher)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
PropertyAOData *pr_ao_data = obj_watcher;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusPropertyAO *pr_ao = pr_ao_data->parent;
nm_assert(g_hash_table_lookup(pr_ao->hash, pr_ao_data) == pr_ao_data);
if (!pr_ao_data->block_is_changed && !pr_ao_data->is_changed) {
pr_ao_data->is_changed = TRUE;
pr_ao_data->changed_next = pr_ao->changed_head;
pr_ao->changed_head = pr_ao_data;
if (!pr_ao->is_changed) {
pr_ao->is_changed = TRUE;
nml_dbus_object_obj_changed_link(self,
pr_ao->owner_dbobj,
NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
}
}
_ASSERT_pr_ao(pr_ao);
}
libnm-utils: expose nml_dbus_property_ao_notify() Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net>
2021-10-07 17:29:13 +02:00
NMLDBusNotifyUpdatePropFlags
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nml_dbus_property_ao_notify(NMClient *self,
NMLDBusPropertyAO *pr_ao,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface *meta_iface,
guint dbus_property_idx,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *value)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
CList stale_lst_head = C_LIST_INIT(stale_lst_head);
PropertyAOData *pr_ao_data;
gboolean changed_prop = FALSE;
gboolean changed_obj = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_ao->owner_dbobj) {
nm_assert(!pr_ao->data_lst_head.next);
nm_assert(!pr_ao->data_lst_head.prev);
nm_assert(!pr_ao->hash);
nm_assert(!pr_ao->meta_iface);
nm_assert(pr_ao->dbus_property_idx == 0);
nm_assert(pr_ao->n_not_ready == 0);
nm_assert(!pr_ao->changed_head);
nm_assert(!pr_ao->is_changed);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_init(&pr_ao->data_lst_head);
pr_ao->hash = g_hash_table_new(nm_ppdirect_hash, nm_ppdirect_equal);
pr_ao->owner_dbobj = dbobj;
pr_ao->meta_iface = meta_iface;
pr_ao->dbus_property_idx = dbus_property_idx;
} else {
nm_assert(pr_ao->data_lst_head.next);
nm_assert(pr_ao->data_lst_head.prev);
nm_assert(pr_ao->hash);
nm_assert(pr_ao->meta_iface == meta_iface);
nm_assert(pr_ao->dbus_property_idx == dbus_property_idx);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_splice(&stale_lst_head, &pr_ao->data_lst_head);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (value) {
GVariantIter iter;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *path;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_iter_init(&iter, value);
while (g_variant_iter_next(&iter, "&o", &path)) {
nm_auto_ref_string NMRefString *dbus_path_r = NULL;
gpointer p_dbus_path_1;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
G_STATIC_ASSERT_EXPR(G_STRUCT_OFFSET(PropertyAOData, obj_watcher) == 0);
G_STATIC_ASSERT_EXPR(G_STRUCT_OFFSET(NMLDBusObjWatcher, dbobj) == 0);
G_STATIC_ASSERT_EXPR(G_STRUCT_OFFSET(NMLDBusObject, dbus_path) == 0);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!nm_dbus_path_not_empty(path)) {
/* should not happen. Anyway, silently skip empty paths. */
continue;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_path_r = nm_ref_string_new(path);
p_dbus_path_1 = &dbus_path_r;
pr_ao_data = g_hash_table_lookup(pr_ao->hash, &p_dbus_path_1);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (pr_ao_data) {
/* With this implementation we cannot track the same path multiple times.
* Of course, for none of the properties where we use this, the server
* should expose the same path more than once, so this limitation is fine
* (maybe even preferable to drop duplicates form NMClient's API). */
nm_assert(pr_ao_data->obj_watcher.dbobj->dbus_path == dbus_path_r);
if (!changed_prop && pr_ao_data->is_notified) {
/* The order of a notified entry changed. That means, we need to signal
* a change of the property. This detection of a change is not always
* correct, in particular we might detect some changes when there were
* none. That's not a serious problem, and fixing it would be expensive
* to implement. */
changed_prop = (c_list_first(&stale_lst_head) != &pr_ao_data->data_lst);
}
nm_c_list_move_tail(&pr_ao->data_lst_head, &pr_ao_data->data_lst);
} else {
pr_ao_data = _dbobjs_obj_watcher_register_r(self,
g_steal_pointer(&dbus_path_r),
nml_dbus_property_ao_notify_watch_cb,
sizeof(PropertyAOData)),
pr_ao_data->parent = pr_ao;
pr_ao_data->nmobj = NULL;
pr_ao_data->changed_next = NULL;
pr_ao_data->is_changed = TRUE;
pr_ao_data->block_is_changed = FALSE;
pr_ao_data->is_ready = FALSE;
pr_ao_data->is_notified = FALSE;
c_list_link_tail(&pr_ao->data_lst_head, &pr_ao_data->data_lst);
if (!g_hash_table_add(pr_ao->hash, pr_ao_data))
nm_assert_not_reached();
nm_assert(pr_ao->n_not_ready < G_MAXUINT);
pr_ao->n_not_ready++;
}
#if NM_MORE_ASSERTS > 10
{
nm_auto_ref_string NMRefString *p = nm_ref_string_new(path);
gpointer pp = &p;
nm_assert(g_hash_table_lookup(pr_ao->hash, &pp) == pr_ao_data);
}
#endif
}
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
pr_ao->changed_head = NULL;
c_list_for_each_entry (pr_ao_data, &pr_ao->data_lst_head, data_lst) {
if (pr_ao_data->is_changed) {
pr_ao_data->changed_next = pr_ao->changed_head;
pr_ao->changed_head = pr_ao_data;
changed_obj = TRUE;
}
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
while ((pr_ao_data = c_list_first_entry(&stale_lst_head, PropertyAOData, data_lst))) {
changed_obj = TRUE;
c_list_unlink(&pr_ao_data->data_lst);
if (!g_hash_table_remove(pr_ao->hash, pr_ao_data))
nm_assert_not_reached();
if (!pr_ao_data->is_ready) {
nm_assert(pr_ao->n_not_ready > 0);
pr_ao->n_not_ready--;
} else {
if (nml_dbus_property_ao_notify_changed_ao(pr_ao_data, self, FALSE))
changed_prop = TRUE;
}
_dbobjs_obj_watcher_unregister(self, pr_ao_data);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_ASSERT_pr_ao(pr_ao);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed_obj) {
pr_ao->is_changed = TRUE;
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed_prop) {
nm_clear_pointer(&pr_ao->arr, g_ptr_array_unref);
_nm_client_queue_notify_object(
self,
pr_ao->owner_dbobj->nmobj,
pr_ao->meta_iface->obj_properties
[pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx].obj_properties_idx]);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NONE;
}
libnm-utils: let nml_dbus_property_ao_clear() return changed_prop Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net>
2021-10-07 18:27:24 +02:00
gboolean
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nml_dbus_property_ao_clear(NMLDBusPropertyAO *pr_ao, NMClient *self)
{
libnm-utils: let nml_dbus_property_ao_clear() return changed_prop Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net>
2021-10-07 18:27:24 +02:00
gboolean changed_prop = FALSE;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_ASSERT_pr_ao(pr_ao);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!pr_ao->owner_dbobj) {
nm_assert(pr_ao->n_not_ready == 0);
nm_assert((!pr_ao->data_lst_head.next && !pr_ao->data_lst_head.prev)
|| (pr_ao->data_lst_head.next == pr_ao->data_lst_head.prev));
nm_assert(!pr_ao->hash);
nm_assert(!pr_ao->meta_iface);
nm_assert(pr_ao->dbus_property_idx == 0);
nm_assert(!pr_ao->is_changed);
} else {
PropertyAOData *pr_ao_data;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
nm_assert(pr_ao->data_lst_head.next);
nm_assert(pr_ao->data_lst_head.prev);
nm_assert(pr_ao->hash);
nm_assert(pr_ao->meta_iface);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
while ((pr_ao_data = c_list_first_entry(&pr_ao->data_lst_head, PropertyAOData, data_lst))) {
if (!pr_ao_data->is_ready) {
nm_assert(pr_ao->n_not_ready > 0);
pr_ao->n_not_ready--;
} else {
if (nml_dbus_property_ao_notify_changed_ao(pr_ao_data, self, FALSE))
changed_prop = TRUE;
}
c_list_unlink(&pr_ao_data->data_lst);
if (!g_hash_table_remove(pr_ao->hash, pr_ao_data))
nm_assert_not_reached();
_dbobjs_obj_watcher_unregister(self, pr_ao_data);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(c_list_is_empty(&pr_ao->data_lst_head));
nm_assert(pr_ao->n_not_ready == 0);
nm_assert(g_hash_table_size(pr_ao->hash) == 0);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed_prop && pr_ao->owner_dbobj->nmobj) {
_nm_client_queue_notify_object(
self,
pr_ao->owner_dbobj->nmobj,
pr_ao->meta_iface
->obj_properties[pr_ao->meta_iface->dbus_properties[pr_ao->dbus_property_idx]
.obj_properties_idx]);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(c_list_is_empty(&pr_ao->data_lst_head));
nm_assert(pr_ao->n_not_ready == 0);
nm_assert(g_hash_table_size(pr_ao->hash) == 0);
nm_clear_pointer(&pr_ao->hash, g_hash_table_unref);
pr_ao->owner_dbobj = NULL;
pr_ao->meta_iface = NULL;
pr_ao->dbus_property_idx = 0;
pr_ao->data_lst_head.next = NULL;
pr_ao->data_lst_head.prev = NULL;
pr_ao->is_changed = FALSE;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_pointer(&pr_ao->arr, g_ptr_array_unref);
libnm-utils: let nml_dbus_property_ao_clear() return changed_prop Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net>
2021-10-07 18:27:24 +02:00
return changed_prop;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
void
nml_dbus_property_ao_clear_many(NMLDBusPropertyAO *pr_ao, guint len, NMClient *self)
{
while (len-- > 0)
nml_dbus_property_ao_clear(pr_ao++, self);
}
/*****************************************************************************/
NMLDBusNotifyUpdatePropFlags
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nml_dbus_notify_update_prop_ignore(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface *meta_iface,
guint dbus_property_idx,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *value)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
return NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NONE;
}
NMLDBusNotifyUpdatePropFlags
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nml_dbus_notify_update_prop_o(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface *meta_iface,
guint dbus_property_idx,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *value)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *path = NULL;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMRefString **p_property;
if (value)
path = g_variant_get_string(value, NULL);
p_property =
nml_dbus_object_get_property_location(dbobj,
meta_iface,
&meta_iface->dbus_properties[dbus_property_idx]);
if (!nm_streq0(nm_ref_string_get_str(*p_property), path)) {
nm_ref_string_unref(*p_property);
*p_property = nm_ref_string_new(path);
}
return NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NOTIFY;
}
/*****************************************************************************/
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_obj_handle_dbus_prop_changes(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObjIfaceData *db_iface_data,
guint dbus_property_idx,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *value)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const NMLDBusMetaIface *meta_iface = db_iface_data->dbus_iface.meta;
const NMLDBusMetaProperty *meta_property = &meta_iface->dbus_properties[dbus_property_idx];
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer p_property;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *dbus_type_s;
const GParamSpec *param_spec;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusNotifyUpdatePropFlags notify_update_prop_flags;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(G_IS_OBJECT(dbobj->nmobj));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (value && !g_variant_is_of_type(value, meta_property->dbus_type)) {
NML_NMCLIENT_LOG_E(self,
"[%s] property %s.%s expected of type \"%s\" but is \"%s\". Ignore",
dbobj->dbus_path->str,
meta_iface->dbus_iface_name,
meta_property->dbus_property_name,
(const char *) meta_property->dbus_type,
(const char *) g_variant_get_type(value));
value = NULL;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
dbus-metadata: make 'extra' available from NML_DBUS_META_PROPERTY_INIT_FCN() Currently a NML_DBUS_META_PROPERTY_INIT_FCN() property does not have 'extra' field available. In order to be able to call 'nml_dbus_property_ao_notify()' from the callback, the 'extra' field must be available. The patch is also dropping 'use_notify_update_prop' field as it only existed to differentiate the union. Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net>
2021-10-07 16:06:14 +02:00
if (meta_property->notify_update_prop) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
notify_update_prop_flags =
meta_property->notify_update_prop(self, dbobj, meta_iface, dbus_property_idx, value);
if (notify_update_prop_flags == NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NONE)
return;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(notify_update_prop_flags == NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NOTIFY);
nm_assert(G_IS_OBJECT(dbobj->nmobj));
nm_assert(meta_iface->obj_properties);
nm_assert(meta_property->obj_properties_idx > 0);
param_spec = meta_iface->obj_properties[meta_property->obj_properties_idx];
goto notify;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
p_property = nml_dbus_object_get_property_location(dbobj, meta_iface, meta_property);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_type_s = (const char *) meta_property->dbus_type;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(G_IS_OBJECT(dbobj->nmobj));
nm_assert(meta_iface->obj_properties);
nm_assert(meta_property->obj_properties_idx > 0);
param_spec = meta_iface->obj_properties[meta_property->obj_properties_idx];
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
notify_update_prop_flags = NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NOTIFY;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
switch (dbus_type_s[0]) {
case 'b':
nm_assert(dbus_type_s[1] == '\0');
if (value)
*((bool *) p_property) = g_variant_get_boolean(value);
else if (param_spec->value_type == G_TYPE_BOOLEAN)
*((bool *) p_property) = ((GParamSpecBoolean *) param_spec)->default_value;
else {
nm_assert_not_reached();
*((bool *) p_property) = FALSE;
}
break;
case 'y':
nm_assert(dbus_type_s[1] == '\0');
if (value)
*((guint8 *) p_property) = g_variant_get_byte(value);
else {
nm_assert(nm_utils_g_param_spec_is_default(param_spec));
*((guint8 *) p_property) = 0;
}
break;
case 'q':
nm_assert(dbus_type_s[1] == '\0');
if (value)
*((guint16 *) p_property) = g_variant_get_uint16(value);
else {
nm_assert(nm_utils_g_param_spec_is_default(param_spec));
*((guint16 *) p_property) = 0;
}
break;
case 'i':
nm_assert(dbus_type_s[1] == '\0');
if (value)
*((gint32 *) p_property) = g_variant_get_int32(value);
else if (param_spec->value_type == G_TYPE_INT)
*((gint32 *) p_property) = ((GParamSpecInt *) param_spec)->default_value;
else {
nm_assert(nm_utils_g_param_spec_is_default(param_spec));
*((gint32 *) p_property) = 0;
}
break;
case 'u':
nm_assert(dbus_type_s[1] == '\0');
if (value)
*((guint32 *) p_property) = g_variant_get_uint32(value);
else {
nm_assert(nm_utils_g_param_spec_is_default(param_spec));
*((guint32 *) p_property) = 0;
}
break;
case 'x':
nm_assert(dbus_type_s[1] == '\0');
if (value)
*((gint64 *) p_property) = g_variant_get_int64(value);
else if (param_spec->value_type == G_TYPE_INT64)
*((gint64 *) p_property) = ((GParamSpecInt64 *) param_spec)->default_value;
else {
nm_assert(nm_utils_g_param_spec_is_default(param_spec));
*((gint64 *) p_property) = 0;
}
break;
case 't':
nm_assert(dbus_type_s[1] == '\0');
if (value)
*((guint64 *) p_property) = g_variant_get_uint64(value);
else {
nm_assert(nm_utils_g_param_spec_is_default(param_spec));
*((guint64 *) p_property) = 0;
}
break;
case 's':
nm_assert(dbus_type_s[1] == '\0');
nm_clear_g_free((char **) p_property);
if (value)
*((char **) p_property) = g_variant_dup_string(value, NULL);
else {
nm_assert(nm_utils_g_param_spec_is_default(param_spec));
*((char **) p_property) = NULL;
}
break;
case 'o':
nm_assert(dbus_type_s[1] == '\0');
notify_update_prop_flags = nml_dbus_property_o_notify(self,
p_property,
dbobj,
meta_iface,
dbus_property_idx,
value);
break;
case 'a':
switch (dbus_type_s[1]) {
case 'y':
nm_assert(dbus_type_s[2] == '\0');
{
gconstpointer v;
gsize l;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GBytes *b = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (value) {
v = g_variant_get_fixed_array(value, &l, 1);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (l > 0) {
/* empty arrays are coerced to NULL. */
b = g_bytes_new(v, l);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_pointer((GBytes **) p_property, g_bytes_unref);
*((GBytes **) p_property) = b;
}
break;
case 's':
nm_assert(dbus_type_s[2] == '\0');
nm_assert(param_spec->value_type == G_TYPE_STRV);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_strfreev(*((char ***) p_property));
if (value)
*((char ***) p_property) = g_variant_dup_strv(value, NULL);
else
*((char ***) p_property) = NULL;
break;
case 'o':
nm_assert(dbus_type_s[2] == '\0');
notify_update_prop_flags = nml_dbus_property_ao_notify(self,
p_property,
dbobj,
meta_iface,
dbus_property_idx,
value);
break;
default:
nm_assert_not_reached();
}
break;
default:
nm_assert_not_reached();
}
notify:
if (NM_FLAGS_HAS(notify_update_prop_flags, NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NOTIFY))
g_object_notify_by_pspec(dbobj->nmobj, (GParamSpec *) param_spec);
}
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_obj_handle_dbus_iface_changes(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObjIfaceData *db_iface_data)
{
NMLDBusObjPropData *db_prop_data;
gboolean is_self = (G_OBJECT(self) == dbobj->nmobj);
gboolean is_removed;
gboolean type_compatible;
guint8 i_prop;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
nm_assert(is_self || NM_IS_OBJECT(dbobj->nmobj));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (G_UNLIKELY(!db_iface_data->nmobj_checked)) {
db_iface_data->nmobj_checked = TRUE;
type_compatible = db_iface_data->dbus_iface.meta->get_type_fcn
&& g_type_is_a(G_OBJECT_TYPE(dbobj->nmobj),
db_iface_data->dbus_iface.meta->get_type_fcn());
db_iface_data->nmobj_compatible = type_compatible;
} else
type_compatible = db_iface_data->nmobj_compatible;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!type_compatible) {
/* on D-Bus, we have this interface associate with the object, but apparently
* it is not compatible. This is either a bug, or NetworkManager exposed an
* unexpected interface on D-Bus object for which we create a certain NMObject
* type. */
return;
}
is_removed = c_list_is_empty(&db_iface_data->iface_lst);
nm_assert(is_removed || !c_list_is_empty(&db_iface_data->changed_prop_lst_head));
_nm_client_queue_notify_object(self, dbobj->nmobj, NULL);
if (is_removed) {
for (i_prop = 0; i_prop < db_iface_data->dbus_iface.meta->n_dbus_properties; i_prop++) {
libnm/client: don't reset properties when interface goes away In case the D-Bus interfaces start dropping off (typically all off them go one by one when the object is being deleted), don't reset all the properties. In particular, keep most properties around, only tear down "o" and "ao", so that the object dependencies get torn down, but we still get enough properties around to identify what the dead object was its heyday. One example of where this is not good is when the device-removed signal is emmitted, the device no longer has the ifname: $ nmcli monitor <quit NetworkManager> (null): device removed (null): device removed ...
2022-06-24 13:59:31 +02:00
const GVariantType *dbus_type =
db_iface_data->dbus_iface.meta->dbus_properties[i_prop].dbus_type;
/* Unset properties that can potentially contain objects, to release them,
* but keep the rest around, because it might still make sense to know what
* they were (e.g. when a device has been removed we'd like know what interface
* name it had, or keep the state to avoid spurious state change into UNKNOWN). */
if (g_variant_type_is_array(dbus_type)
|| g_variant_type_equal(dbus_type, G_VARIANT_TYPE_OBJECT_PATH))
_obj_handle_dbus_prop_changes(self, dbobj, db_iface_data, i_prop, NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
} else {
while ((db_prop_data = c_list_first_entry(&db_iface_data->changed_prop_lst_head,
NMLDBusObjPropData,
changed_prop_lst))) {
gs_unref_variant GVariant *prop_data_value = NULL;
c_list_unlink(&db_prop_data->changed_prop_lst);
nm_assert(db_prop_data >= db_iface_data->prop_datas);
nm_assert(
db_prop_data
< &db_iface_data->prop_datas[db_iface_data->dbus_iface.meta->n_dbus_properties]);
nm_assert(db_prop_data->prop_data_value);
all: fix typo in man pages There should be a comma after 'Otherwise' and 'Currently'. https://bugzilla.redhat.com/show_bug.cgi?id=1852452 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/560
2020-07-01 17:20:40 -04:00
/* Currently, NMLDBusObject forgets about the variant. Theoretically, it could cache
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
* it, but there is no need because we update the property in nmobj (which extracts and
* keeps the property value itself).
*
* Note that we only consume the variant here when we process it.
* That implies that we already created a NMObject for the dbobj
* instance. Unless that happens, we cache the last seen property values. */
prop_data_value = g_steal_pointer(&db_prop_data->prop_data_value);
i_prop = (db_prop_data - &db_iface_data->prop_datas[0]);
_obj_handle_dbus_prop_changes(self, dbobj, db_iface_data, i_prop, prop_data_value);
}
}
}
static void
_obj_handle_dbus_changes(NMClient *self, NMLDBusObject *dbobj)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
NMLDBusObjIfaceData *db_iface_data;
NMLDBusObjIfaceData *db_iface_data_safe;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_object GObject *nmobj_unregistering = NULL;
_ASSERT_dbobj(dbobj, self);
all: fix minor typos https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/565
2020-07-04 11:37:01 +03:00
/* In a first step we only remember all the changes that a D-Bus message brings
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
* and queue the object to process them.
*
* Here (in step 2) we look at what changed on D-Bus and propagate those changes
* to the NMObject instance.
*
* Note that here we still must not emit any GObject signals. That follows later,
* and again if the object changes, we will just queue that we handle the changes
* later. */
c_list_for_each_entry_safe (db_iface_data,
db_iface_data_safe,
&dbobj->iface_lst_head,
iface_lst) {
if (!db_iface_data->iface_removed)
continue;
c_list_unlink(&db_iface_data->iface_lst);
if (db_iface_data->dbus_iface_is_wellknown && dbobj->nmobj)
_obj_handle_dbus_iface_changes(self, dbobj, db_iface_data);
nml_dbus_obj_iface_data_destroy(db_iface_data);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (G_UNLIKELY(!dbobj->nmobj) && !c_list_is_empty(&dbobj->iface_lst_head)) {
/* Try to create a NMObject for this D-Bus object. Note that we detect the type
* based on the interfaces that it has, and if we make a choice once, we don't
* change. That means, one D-Bus object can only be of one type. */
if (NM_IN_SET(dbobj->dbus_path,
_dbus_path_nm,
_dbus_path_settings,
_dbus_path_dns_manager)) {
/* For the main types, we don't detect them based on the interfaces present,
* but on the path names. Of course, both should correspond anyway. */
NML_NMCLIENT_LOG_T(self,
"[%s]: register NMClient for D-Bus object",
dbobj->dbus_path->str);
dbobj->nmobj = G_OBJECT(self);
if (dbobj->dbus_path == _dbus_path_nm) {
nm_assert(!priv->dbobj_nm);
priv->dbobj_nm = dbobj;
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
_set_nm_running(self, TRUE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
} else if (dbobj->dbus_path == _dbus_path_settings) {
nm_assert(!priv->dbobj_settings);
priv->dbobj_settings = dbobj;
} else {
nm_assert(dbobj->dbus_path == _dbus_path_dns_manager);
nm_assert(!priv->dbobj_dns_manager);
priv->dbobj_dns_manager = dbobj;
}
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY, self);
} else {
GType gtype = G_TYPE_NONE;
libnm/trivial: rename enums NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_* We will need more levels of priority. Change the naming to make room for that. sed 's/NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_LOW/NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_10/g' `git grep -l NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_ ` -i sed 's/NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_LOW/NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_10/g' `git grep -l NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_ ` -i ./contrib/scripts/nm-code-format-container.sh Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net> [thaller@redhat.com: split original patch]
2020-11-18 10:16:40 +01:00
NMLDBusMetaInteracePrio curr_prio = NML_DBUS_META_INTERFACE_PRIO_INSTANTIATE_10 - 1;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_for_each_entry (db_iface_data, &dbobj->iface_lst_head, iface_lst) {
nm_assert(!db_iface_data->iface_removed);
if (!db_iface_data->dbus_iface_is_wellknown)
break;
if (db_iface_data->dbus_iface.meta->interface_prio <= curr_prio)
continue;
curr_prio = db_iface_data->dbus_iface.meta->interface_prio;
gtype = db_iface_data->dbus_iface.meta->get_type_fcn();
}
if (gtype != G_TYPE_NONE) {
dbobj->nmobj = g_object_new(gtype, NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(self,
"[%s]: register new NMObject " NM_HASH_OBFUSCATE_PTR_FMT
" of type %s",
dbobj->dbus_path->str,
NM_HASH_OBFUSCATE_PTR(dbobj->nmobj),
g_type_name(gtype));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_OBJECT(dbobj->nmobj));
NM_OBJECT_GET_CLASS(dbobj->nmobj)
->register_client(NM_OBJECT(dbobj->nmobj), self, dbobj);
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY, self);
}
}
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_for_each_entry (db_iface_data, &dbobj->iface_lst_head, iface_lst) {
nm_assert(!db_iface_data->iface_removed);
if (!db_iface_data->dbus_iface_is_wellknown)
break;
if (c_list_is_empty(&db_iface_data->changed_prop_lst_head))
continue;
if (dbobj->nmobj)
_obj_handle_dbus_iface_changes(self, dbobj, db_iface_data);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (c_list_is_empty(&dbobj->iface_lst_head) && dbobj->nmobj) {
if (dbobj->nmobj == G_OBJECT(self)) {
dbobj->nmobj = NULL;
NML_NMCLIENT_LOG_T(self,
"[%s]: unregister NMClient from D-Bus object",
dbobj->dbus_path->str);
if (dbobj->dbus_path == _dbus_path_nm) {
nm_assert(priv->dbobj_nm == dbobj);
priv->dbobj_nm = NULL;
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
_set_nm_running(self, TRUE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nml_dbus_property_o_clear_many(priv->nm.property_o,
G_N_ELEMENTS(priv->nm.property_o),
self);
nml_dbus_property_ao_clear_many(priv->nm.property_ao,
G_N_ELEMENTS(priv->nm.property_ao),
self);
} else if (dbobj->dbus_path == _dbus_path_settings) {
nm_assert(priv->dbobj_settings == dbobj);
priv->dbobj_settings = NULL;
nml_dbus_property_ao_clear(&priv->settings.connections, self);
} else {
nm_assert(dbobj->dbus_path == _dbus_path_dns_manager);
nm_assert(priv->dbobj_dns_manager == dbobj);
priv->dbobj_dns_manager = NULL;
}
} else {
nmobj_unregistering = g_steal_pointer(&dbobj->nmobj);
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_WATCHED_ONLY, self);
NML_NMCLIENT_LOG_T(self,
"[%s]: unregister NMObject " NM_HASH_OBFUSCATE_PTR_FMT " of type %s",
dbobj->dbus_path->str,
NM_HASH_OBFUSCATE_PTR(nmobj_unregistering),
g_type_name(G_OBJECT_TYPE(nmobj_unregistering)));
NM_OBJECT_GET_CLASS(nmobj_unregistering)
->unregister_client(NM_OBJECT(nmobj_unregistering), self, dbobj);
}
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ);
}
/*****************************************************************************/
static void
_dbus_handle_obj_changed_nmobj(NMClient *self)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
CList obj_changed_tmp_lst_head = C_LIST_INIT(obj_changed_tmp_lst_head);
nm_assert(!nml_dbus_object_obj_changed_any_linked(self, ~NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ));
/* First we notify all watchers that these objects changed. Note that we only do that
* here for the list before processing the changes below in a loop. That is, because
* processing changes can again enqueue changed objects, and we only want to want to
* notify watchers for the events that happened earlier (not repeatedly notify them). */
c_list_splice(&obj_changed_tmp_lst_head, &priv->obj_changed_lst_head);
while (
(dbobj = c_list_first_entry(&obj_changed_tmp_lst_head, NMLDBusObject, obj_changed_lst))) {
nm_c_list_move_tail(&priv->obj_changed_lst_head, &dbobj->obj_changed_lst);
_dbobjs_notify_watchers_for_dbobj(self, dbobj);
}
again:
nm_assert(!nml_dbus_object_obj_changed_any_linked(self, ~NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_splice(&obj_changed_tmp_lst_head, &priv->obj_changed_lst_head);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
while (
(dbobj = c_list_first_entry(&obj_changed_tmp_lst_head, NMLDBusObject, obj_changed_lst))) {
if (!nml_dbus_object_obj_changed_consume(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ)) {
nm_assert_not_reached();
continue;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!dbobj->nmobj)
continue;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (dbobj->nmobj == G_OBJECT(self)) {
if (dbobj == priv->dbobj_nm) {
nml_dbus_property_o_notify_changed_many(priv->nm.property_o,
G_N_ELEMENTS(priv->nm.property_o),
self);
nml_dbus_property_ao_notify_changed_many(priv->nm.property_ao,
G_N_ELEMENTS(priv->nm.property_ao),
self);
} else if (dbobj == priv->dbobj_settings)
nml_dbus_property_ao_notify_changed(&priv->settings.connections, self);
else
nm_assert(dbobj == priv->dbobj_dns_manager);
} else
NM_OBJECT_GET_CLASS(dbobj->nmobj)->obj_changed_notify(NM_OBJECT(dbobj->nmobj));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_dbobjs_check_dbobj_ready(self, dbobj);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!c_list_is_empty(&priv->obj_changed_lst_head)) {
nm_assert(nml_dbus_object_obj_changed_any_linked(self, NML_DBUS_OBJ_CHANGED_TYPE_NMOBJ));
/* we got new changes enqueued. Need to check again. */
goto again;
}
}
static void
_dbus_handle_obj_changed_dbus(NMClient *self, const char *log_context)
{
NMClientPrivate *priv;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
CList obj_changed_tmp_lst_head = C_LIST_INIT(obj_changed_tmp_lst_head);
priv = NM_CLIENT_GET_PRIVATE(self);
/* We move the changed list onto a temporary list and consume that.
* Note that nml_dbus_object_obj_changed_consume() will move the object
* back to the original list if there are changes of another type.
*
* This is done so that we can enqueue more changes while processing the
* change list. */
c_list_splice(&obj_changed_tmp_lst_head, &priv->obj_changed_lst_head);
while (
(dbobj = c_list_first_entry(&obj_changed_tmp_lst_head, NMLDBusObject, obj_changed_lst))) {
nm_auto_unref_nml_dbusobj NMLDBusObject *dbobj_unref = NULL;
if (!nml_dbus_object_obj_changed_consume(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_DBUS))
continue;
nm_assert(dbobj->obj_state >= NML_DBUS_OBJ_STATE_ON_DBUS);
dbobj_unref = nml_dbus_object_ref(dbobj);
_obj_handle_dbus_changes(self, dbobj);
if (dbobj->obj_state == NML_DBUS_OBJ_STATE_UNLINKED)
continue;
if (c_list_is_empty(&dbobj->iface_lst_head) && c_list_is_empty(&dbobj->watcher_lst_head)) {
NML_NMCLIENT_LOG_T(self, "[%s]: drop D-Bus instance", dbobj->dbus_path->str);
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_UNLINKED, self);
if (!g_hash_table_steal(priv->dbus_objects, dbobj))
nm_assert_not_reached();
nml_dbus_object_unref(dbobj);
}
}
/* D-Bus changes can only be enqueued in an earlier stage. We don't expect
* anymore changes of type D-Bus at this point. */
nm_assert(!nml_dbus_object_obj_changed_any_linked(self, NML_DBUS_OBJ_CHANGED_TYPE_DBUS));
}
static void
_dbus_handle_changes_commit(NMClient *self, gboolean allow_init_start_check_complete)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_auto_pop_gmaincontext GMainContext *dbus_context = NULL;
_dbus_handle_obj_changed_nmobj(self);
dbus_context = nm_g_main_context_push_thread_default_if_necessary(priv->main_context);
_nm_client_notify_event_emit(self);
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
_set_nm_running(self, FALSE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (allow_init_start_check_complete)
_init_start_check_complete(self);
}
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_dbus_handle_changes(NMClient *self,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const char *log_context,
gboolean allow_init_start_check_complete)
{
_dbus_handle_obj_changed_dbus(self, log_context);
_dbus_handle_changes_commit(self, allow_init_start_check_complete);
}
static gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_dbus_handle_properties_changed(NMClient *self,
const char *log_context,
const char *object_path,
const char *interface_name,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean allow_add_iface,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *changed_properties,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMLDBusObject **inout_dbobj)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj = NULL;
NMLDBusObjIfaceData *db_iface_data = NULL;
nm_auto_ref_string NMRefString *dbus_path = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(!changed_properties
|| g_variant_is_of_type(changed_properties, G_VARIANT_TYPE("a{sv}")));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
gs_free char *ss = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(self,
libnm: avoid assertion failure in _dbus_handle_properties_changed() for logging no properties (cherry picked from commit 1f9622358a691aa97b72b3df118feac7970769c4)
2021-02-15 09:11:42 +01:00
"[%s]: %s: properties changed for interface %s %s%s%s",
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
object_path,
log_context,
interface_name,
libnm: avoid assertion failure in _dbus_handle_properties_changed() for logging no properties (cherry picked from commit 1f9622358a691aa97b72b3df118feac7970769c4)
2021-02-15 09:11:42 +01:00
NM_PRINT_FMT_QUOTED(changed_properties,
"{ ",
(ss = g_variant_print(changed_properties, TRUE)),
" }",
"(no changed properties)"));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (inout_dbobj) {
dbobj = *inout_dbobj;
nm_assert(!dbobj || nm_streq(object_path, dbobj->dbus_path->str));
}
if (!dbobj) {
dbus_path = nm_ref_string_new(object_path);
dbobj = _dbobjs_dbobj_get_r(self, dbus_path);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: fix tracking object state in NMClient cache NMClient has a NMLDBusObject instance for each D-Bus object that it sees. This object can be in different states, like that we already saw it on D-Bus or that it is only referred to by another property. Due to a bug, we would wrongly not update the state and trigger an assertion. Reproduce with python-dbusmock (commit e89e28bf1bc0254a1eb71b71cf68ef7a97d11e5b) by running `pytest -v -s tests/test_networkmanager.py -k test_one_wifi_with_accesspoints`. With LIBNM_CLIENT_DEBUG we get: >>> libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: properties changed for interface org.freedesktop.NetworkManager.Device { {'ActiveConnection': <objectpath '/org/freedesktop/NetworkManager/ActiveConnection/0'>} } libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: set property org.freedesktop.NetworkManager.Device.ActiveConnection libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 consumed >>> libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: set D-Bus object state watched-only libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 consumed >>> libnm-dbus[96085]: <error> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: property ActiveConnection references /org/freedesktop/NetworkManager/ActiveConnection/0 but object is not present on D-Bus libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: properties changed for interface org.freedesktop.NetworkManager.Device { {'State': <uint32 100>} } libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: set property org.freedesktop.NetworkManager.Device.State libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 consumed libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06461] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06461] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: properties changed for interface org.freedesktop.NetworkManager.Device { {'StateReason': <(uint32 100, uint32 0)>} } libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: set property org.freedesktop.NetworkManager.Device.StateReason libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 consumed libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 consumed >>> libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: properties changed for interface org.freedesktop.NetworkManager.Connection.Active { {'Devices': <[objectpath '/org/freedesktop/NetworkManager/Devices/mock_WiFi2']>, 'Default6': <false>, 'Default': <true>, 'Type': <'802-11-wireless'>, 'Vpn': <false>, 'Connection': <objectpath '/org/freedesktop/NetworkManager/Settings/Mock_AP3'>, 'Master': <objectpath '/'>, 'SpecificObject': <objectpath '/org/freedesktop/NetworkManager/AccessPoint/Mock_AP3'>, 'Uuid': <'72757a57-8cb6-4052-a18f-4e2be4ba27d9'>, 'State': <uint32 2>, 'Id': <'AP_3'>} } >>> here we lack "set D-Bus object state on-dbus" libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Devices libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Default6 libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Default libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Type libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Vpn libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Connection libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Master libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.SpecificObject libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Uuid libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.State libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Id libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: changed-type 0x01 consumed Bail out! libnm:ERROR:libnm/nm-client.c:2863:_dbus_handle_obj_changed_dbus: assertion failed: (dbobj->obj_state >= NML_DBUS_OBJ_STATE_ON_DBUS) Backtrace: #3 0x00007f0bd11173bf in g_assertion_message_expr (domain=domain@entry=0x7f0bd1576018 "libnm", file=file@entry=0x7f0bd1576006 "libnm/nm-client.c", line=line@entry=2863, func=func@entry=0x7f0bd157f1b0 <__func__.170> "_dbus_handle_obj_changed_dbus", expr=expr@entry=0x7f0bd157cba0 "dbobj->obj_state >= NML_DBUS_OBJ_STATE_ON_DBUS") at ../glib/gtestutils.c:2963 #4 0x00007f0bd14959dd in _dbus_handle_obj_changed_dbus (self=self@entry=0x5612d4f5a130, log_context=<optimized out>) at libnm/nm-client.c:2863 #5 0x00007f0bd1495c29 in _dbus_handle_changes (self=self@entry=0x5612d4f5a130, log_context=<optimized out>, allow_init_start_check_complete=allow_init_start_check_complete@entry=1) at libnm/nm-client.c:2909 #6 0x00007f0bd1497e56 in _dbus_managed_objects_changed_cb (connection=<optimized out>, sender_name=<optimized out>, arg_object_path=<optimized out>, interface_name=<optimized out>, signal_name=<optimized out>, parameters=0x7f0bb800d720, user_data=0x5612d4f5a130) at libnm/nm-client.c:3172 #7 0x00007f0bd132a8df in emit_signal_instance_in_idle_cb (data=data@entry=0x7f0bb8003700) at ../gio/gdbusconnection.c:3789 #8 0x00007f0bd10f1b5b in g_idle_dispatch (source=source@entry=0x7f0bb8012260, callback=0x7f0bd132a860 <emit_signal_instance_in_idle_cb>, user_data=0x7f0bb8003700) at ../glib/gmain.c:5836 #9 0x00007f0bd10f2a9f in g_main_dispatch (context=0x5612d4f4b630) at ../glib/gmain.c:3325 #10 g_main_context_dispatch (context=0x5612d4f4b630) at ../glib/gmain.c:4043 #11 0x00007f0bd1144a98 in g_main_context_iterate.constprop.0 (context=0x5612d4f4b630, block=block@entry=1, dispatch=dispatch@entry=1, self=<optimized out>) at ../glib/gmain.c:4119 #12 0x00007f0bd10f2163 in g_main_loop_run (loop=0x5612d4f4b720) at ../glib/gmain.c:4317 #13 0x00005612d44b6543 in main (argc=7, argv=0x7fff4414f1d8) at clients/cli/nmcli.c:1036 https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=982613 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/662 Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') (cherry picked from commit e1e9abdf041b4cc95fb1936b75ced7669f3d7867)
2021-02-15 09:17:07 +01:00
if (dbobj) {
nm_assert(dbobj->obj_state >= NML_DBUS_OBJ_STATE_WATCHED_ONLY);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
db_iface_data = nml_dbus_object_iface_data_get(dbobj, interface_name, allow_add_iface);
libnm: fix tracking object state in NMClient cache NMClient has a NMLDBusObject instance for each D-Bus object that it sees. This object can be in different states, like that we already saw it on D-Bus or that it is only referred to by another property. Due to a bug, we would wrongly not update the state and trigger an assertion. Reproduce with python-dbusmock (commit e89e28bf1bc0254a1eb71b71cf68ef7a97d11e5b) by running `pytest -v -s tests/test_networkmanager.py -k test_one_wifi_with_accesspoints`. With LIBNM_CLIENT_DEBUG we get: >>> libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: properties changed for interface org.freedesktop.NetworkManager.Device { {'ActiveConnection': <objectpath '/org/freedesktop/NetworkManager/ActiveConnection/0'>} } libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: set property org.freedesktop.NetworkManager.Device.ActiveConnection libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 consumed >>> libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: set D-Bus object state watched-only libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 consumed >>> libnm-dbus[96085]: <error> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: property ActiveConnection references /org/freedesktop/NetworkManager/ActiveConnection/0 but object is not present on D-Bus libnm-dbus[96085]: <trace> [6464.06459] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: properties changed for interface org.freedesktop.NetworkManager.Device { {'State': <uint32 100>} } libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: set property org.freedesktop.NetworkManager.Device.State libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 consumed libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06460] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06461] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06461] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: properties changed for interface org.freedesktop.NetworkManager.Device { {'StateReason': <(uint32 100, uint32 0)>} } libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: properties-changed: set property org.freedesktop.NetworkManager.Device.StateReason libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x01 consumed libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 linked libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/Devices/mock_WiFi2]: changed-type 0x02 consumed libnm-dbus[96085]: <trace> [6464.06462] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager]: changed-type 0x02 consumed >>> libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: properties changed for interface org.freedesktop.NetworkManager.Connection.Active { {'Devices': <[objectpath '/org/freedesktop/NetworkManager/Devices/mock_WiFi2']>, 'Default6': <false>, 'Default': <true>, 'Type': <'802-11-wireless'>, 'Vpn': <false>, 'Connection': <objectpath '/org/freedesktop/NetworkManager/Settings/Mock_AP3'>, 'Master': <objectpath '/'>, 'SpecificObject': <objectpath '/org/freedesktop/NetworkManager/AccessPoint/Mock_AP3'>, 'Uuid': <'72757a57-8cb6-4052-a18f-4e2be4ba27d9'>, 'State': <uint32 2>, 'Id': <'AP_3'>} } >>> here we lack "set D-Bus object state on-dbus" libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Devices libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Default6 libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Default libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Type libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Vpn libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Connection libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Master libnm-dbus[96085]: <trace> [6464.06465] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.SpecificObject libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Uuid libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.State libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: interfaces-added: set property org.freedesktop.NetworkManager.Connection.Active.Id libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: changed-type 0x01 linked libnm-dbus[96085]: <trace> [6464.06466] nmclient[c9bf1eaa1f4b6c99]: [/org/freedesktop/NetworkManager/ActiveConnection/0]: changed-type 0x01 consumed Bail out! libnm:ERROR:libnm/nm-client.c:2863:_dbus_handle_obj_changed_dbus: assertion failed: (dbobj->obj_state >= NML_DBUS_OBJ_STATE_ON_DBUS) Backtrace: #3 0x00007f0bd11173bf in g_assertion_message_expr (domain=domain@entry=0x7f0bd1576018 "libnm", file=file@entry=0x7f0bd1576006 "libnm/nm-client.c", line=line@entry=2863, func=func@entry=0x7f0bd157f1b0 <__func__.170> "_dbus_handle_obj_changed_dbus", expr=expr@entry=0x7f0bd157cba0 "dbobj->obj_state >= NML_DBUS_OBJ_STATE_ON_DBUS") at ../glib/gtestutils.c:2963 #4 0x00007f0bd14959dd in _dbus_handle_obj_changed_dbus (self=self@entry=0x5612d4f5a130, log_context=<optimized out>) at libnm/nm-client.c:2863 #5 0x00007f0bd1495c29 in _dbus_handle_changes (self=self@entry=0x5612d4f5a130, log_context=<optimized out>, allow_init_start_check_complete=allow_init_start_check_complete@entry=1) at libnm/nm-client.c:2909 #6 0x00007f0bd1497e56 in _dbus_managed_objects_changed_cb (connection=<optimized out>, sender_name=<optimized out>, arg_object_path=<optimized out>, interface_name=<optimized out>, signal_name=<optimized out>, parameters=0x7f0bb800d720, user_data=0x5612d4f5a130) at libnm/nm-client.c:3172 #7 0x00007f0bd132a8df in emit_signal_instance_in_idle_cb (data=data@entry=0x7f0bb8003700) at ../gio/gdbusconnection.c:3789 #8 0x00007f0bd10f1b5b in g_idle_dispatch (source=source@entry=0x7f0bb8012260, callback=0x7f0bd132a860 <emit_signal_instance_in_idle_cb>, user_data=0x7f0bb8003700) at ../glib/gmain.c:5836 #9 0x00007f0bd10f2a9f in g_main_dispatch (context=0x5612d4f4b630) at ../glib/gmain.c:3325 #10 g_main_context_dispatch (context=0x5612d4f4b630) at ../glib/gmain.c:4043 #11 0x00007f0bd1144a98 in g_main_context_iterate.constprop.0 (context=0x5612d4f4b630, block=block@entry=1, dispatch=dispatch@entry=1, self=<optimized out>) at ../glib/gmain.c:4119 #12 0x00007f0bd10f2163 in g_main_loop_run (loop=0x5612d4f4b720) at ../glib/gmain.c:4317 #13 0x00005612d44b6543 in main (argc=7, argv=0x7fff4414f1d8) at clients/cli/nmcli.c:1036 https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=982613 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/662 Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') (cherry picked from commit e1e9abdf041b4cc95fb1936b75ced7669f3d7867)
2021-02-15 09:17:07 +01:00
if (db_iface_data && dbobj->obj_state == NML_DBUS_OBJ_STATE_WATCHED_ONLY)
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_ON_DBUS, self);
} else if (allow_add_iface) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbobj = _dbobjs_dbobj_create(self, g_steal_pointer(&dbus_path));
nml_dbus_object_set_obj_state(dbobj, NML_DBUS_OBJ_STATE_ON_DBUS, self);
db_iface_data = nml_dbus_object_iface_data_get(dbobj, interface_name, TRUE);
libnm: add assertion in _dbus_handle_properties_changed() (cherry picked from commit 3ceec9c6ac399b849fcabc52cf58e9001a6b8684)
2021-02-15 09:59:53 +01:00
nm_assert(db_iface_data);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NM_SET_OUT(inout_dbobj, dbobj);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!db_iface_data) {
if (allow_add_iface)
NML_NMCLIENT_LOG_E(self,
"%s: [%s] too many interfaces on object. Something is very wrong",
log_context,
object_path);
else
NML_NMCLIENT_LOG_E(self,
"%s: [%s] property changed signal for non existing interface %s",
log_context,
object_path,
interface_name);
nm_assert(!dbobj || dbobj->obj_state != NML_DBUS_OBJ_STATE_UNLINKED);
return FALSE;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!db_iface_data->dbus_iface_is_wellknown)
NML_NMCLIENT_LOG_W(self,
"%s: [%s] ignore unknown interface %s",
log_context,
object_path,
interface_name);
else if (changed_properties) {
GVariantIter iter_prop;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *property_name;
GVariant *property_value_tmp;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_iter_init(&iter_prop, changed_properties);
while (g_variant_iter_next(&iter_prop, "{&sv}", &property_name, &property_value_tmp)) {
_nm_unused gs_unref_variant GVariant *property_value = property_value_tmp;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const NMLDBusMetaProperty *meta_property;
NMLDBusObjPropData *db_propdata;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
guint property_idx;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
meta_property = nml_dbus_meta_property_get(db_iface_data->dbus_iface.meta,
property_name,
&property_idx);
if (!meta_property) {
NML_NMCLIENT_LOG_W(self,
"%s: [%s]: ignore unknown property %s.%s",
log_context,
object_path,
interface_name,
property_name);
continue;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
db_propdata = &db_iface_data->prop_datas[property_idx];
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(self,
"[%s]: %s: %s property %s.%s",
object_path,
log_context,
db_propdata->prop_data_value ? "update" : "set",
interface_name,
property_name);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_g_variant_unref(db_propdata->prop_data_value);
db_propdata->prop_data_value = g_steal_pointer(&property_value);
nm_c_list_move_tail(&db_iface_data->changed_prop_lst_head,
&db_propdata->changed_prop_lst);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_DBUS);
return TRUE;
}
static gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_dbus_handle_interface_added(NMClient *self,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const char *log_context,
const char *object_path,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *ifaces)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
libnm/client: drop an unused argument https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:09:55 +01:00
gboolean changed = FALSE;
const char *interface_name;
GVariant *changed_properties;
GVariantIter iter_ifaces;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(g_variant_is_of_type(ifaces, G_VARIANT_TYPE("a{sa{sv}}")));
g_variant_iter_init(&iter_ifaces, ifaces);
while (g_variant_iter_next(&iter_ifaces, "{&s@a{sv}}", &interface_name, &changed_properties)) {
_nm_unused gs_unref_variant GVariant *changed_properties_free = changed_properties;
if (_dbus_handle_properties_changed(self,
log_context,
object_path,
interface_name,
TRUE,
changed_properties,
libnm/client: drop an unused argument https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:09:55 +01:00
NULL))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
changed = TRUE;
}
return changed;
}
static gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_dbus_handle_interface_removed(NMClient *self,
const char *log_context,
const char *object_path,
NMLDBusObject **inout_dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const char *const *removed_interfaces)
{
gboolean changed = FALSE;
NMLDBusObject *dbobj;
gsize i;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (inout_dbobj && *inout_dbobj) {
dbobj = *inout_dbobj;
nm_assert(dbobj == _dbobjs_dbobj_get_s(self, object_path));
} else {
dbobj = _dbobjs_dbobj_get_s(self, object_path);
if (!dbobj) {
NML_NMCLIENT_LOG_E(self,
"%s: [%s]: receive interface removed event for non existing object",
log_context,
object_path);
return FALSE;
}
NM_SET_OUT(inout_dbobj, dbobj);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
for (i = 0; removed_interfaces[i]; i++) {
NMLDBusObjIfaceData *db_iface_data;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *interface_name = removed_interfaces[i];
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
db_iface_data = nml_dbus_object_iface_data_get(dbobj, interface_name, FALSE);
if (!db_iface_data) {
NML_NMCLIENT_LOG_E(
self,
"%s: [%s] receive interface remove event for unexpected interface %s",
log_context,
object_path,
interface_name);
continue;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(self,
"%s: [%s] receive interface remove event for interface %s",
log_context,
object_path,
interface_name);
db_iface_data->iface_removed = TRUE;
changed = TRUE;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed)
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_DBUS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return changed;
}
static void
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
_dbus_managed_objects_changed_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *arg_object_path,
const char *interface_name,
const char *signal_name,
GVariant *parameters,
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
gpointer user_data)
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *log_context;
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
gboolean changed;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(nm_streq0(interface_name, DBUS_INTERFACE_OBJECT_MANAGER));
if (priv->get_managed_objects_cancellable) {
/* we still wait for the initial GetManagedObjects(). Ignore the event. */
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
return;
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (nm_streq(signal_name, "InterfacesAdded")) {
gs_unref_variant GVariant *interfaces_and_properties = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *object_path;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("(oa{sa{sv}})")))
return;
g_variant_get(parameters, "(&o@a{sa{sv}})", &object_path, &interfaces_and_properties);
log_context = "interfaces-added";
changed =
_dbus_handle_interface_added(self, log_context, object_path, interfaces_and_properties);
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
goto out;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (nm_streq(signal_name, "InterfacesRemoved")) {
gs_free const char **interfaces = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *object_path;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("(oas)")))
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
return;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_get(parameters, "(&o^a&s)", &object_path, &interfaces);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
log_context = "interfaces-removed";
changed = _dbus_handle_interface_removed(self, log_context, object_path, NULL, interfaces);
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
goto out;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
return;
out:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed)
_dbus_handle_changes(self, log_context, TRUE);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static void
_dbus_properties_changed_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *object_path,
const char *signal_interface_name,
const char *signal_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer user_data)
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
const char *interface_name;
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
gs_unref_variant GVariant *changed_properties = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_free const char **invalidated_properties = NULL;
const char *log_context = "properties-changed";
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->get_managed_objects_cancellable) {
/* we still wait for the initial GetManagedObjects(). Ignore the event. */
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
return;
}
libnm: avoid leaking GMainContext for sync initialization after context-busy-watcher quits When passing a destroy notify to g_dbus_connection_signal_subscribe(), that callback gets invoked as an idle handler of the associated GMainContext. That caused to have yet another source attached to the context after the NMClient gets destroyed. Especially with synchronous initialization of NMClient that is bad, because we may destroy the context-busy-watcher too early. That results in removing the integration of the inner GMainContext into the caller's context, and thus we leak the inner context indefinitely. Avoid that leak by not passing a cleanup function to g_dbus_connection_signal_subscribe(). Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-14 11:28:54 +01:00
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("(sa{sv}as)")))
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
return;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_get(parameters,
"(&s@a{sv}^a&s)",
&interface_name,
&changed_properties,
&invalidated_properties);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (invalidated_properties && invalidated_properties[0]) {
NML_NMCLIENT_LOG_W(self,
"%s: [%s] ignore invalidated properties on interface %s",
log_context,
object_path,
libnm: add nm_client_dbus_set_property() API Similar to nm_client_dbus_call(), but useful for setting a D-Bus property on NetworkManager's D-Bus interface. Note that we currently have various synchronous API for setting D-Bus properties (like nm_client_networking_set_enabled()). Synchronous API does not play well with the content of NMClient's cache, and was thus deprecated. However, until now no async variant exists. Instead of adding multiple async operations, I think it should be sufficient to only add one nm_client_dbus_set_property() property. It's still reasonably convenient to use for setting a property.
2020-03-13 18:11:56 +01:00
interface_name);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (_dbus_handle_properties_changed(self,
log_context,
object_path,
interface_name,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
FALSE,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
changed_properties,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
NULL))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_dbus_handle_changes(self, log_context, TRUE);
}
static void
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
_dbus_get_managed_objects_cb(GObject *source, GAsyncResult *result, gpointer user_data)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self;
NMClientPrivate *priv;
gs_unref_variant GVariant *ret = NULL;
gs_unref_variant GVariant *managed_objects = NULL;
gs_free_error GError *error = NULL;
gs_unref_object GObject *context_busy_watcher = NULL;
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
nm_utils_user_data_unpack(user_data, &self, &context_busy_watcher);
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(source), result, &error);
nm_assert((!!ret) != (!!error));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!ret && nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return;
priv = NM_CLIENT_GET_PRIVATE(self);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
if (ret) {
nm_assert(g_variant_is_of_type(ret, G_VARIANT_TYPE("(a{oa{sa{sv}}})")));
managed_objects = g_variant_get_child_value(ret, 0);
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_clear_object(&priv->get_managed_objects_cancellable);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
if (!managed_objects) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "GetManagedObjects() call failed: %s", error->message);
/* hm, now that's odd. Maybe NetworkManager just quit and we are about to get
* a name-owner changed signal soon. Treat this as if we got no managed objects at all. */
} else
NML_NMCLIENT_LOG_D(self, "GetManagedObjects() completed");
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
if (managed_objects) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GVariantIter iter;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *object_path;
GVariant *ifaces_tmp;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
g_variant_iter_init(&iter, managed_objects);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
while (g_variant_iter_next(&iter, "{&o@a{sa{sv}}}", &object_path, &ifaces_tmp)) {
gs_unref_variant GVariant *ifaces = ifaces_tmp;
_dbus_handle_interface_added(self, "get-managed-objects", object_path, ifaces);
}
}
/* always call _dbus_handle_changes(), even if nothing changed. We need this to complete
* initialization. */
_dbus_handle_changes(self, "get-managed-objects", TRUE);
}
/*****************************************************************************/
static void
_nm_client_get_settings_call_cb(GObject *source, GAsyncResult *result, gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMRemoteConnection *remote_connection;
NMClient *self;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_variant GVariant *ret = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_free_error GError *error = NULL;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_variant GVariant *settings = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(source), result, &error);
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!ret && nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return;
remote_connection = user_data;
self = _nm_object_get_client(remote_connection);
dbobj = _nm_object_get_dbobj(remote_connection);
_ASSERT_dbobj(dbobj, self);
if (!ret) {
NML_NMCLIENT_LOG_T(self,
"[%s] GetSettings() completed with error: %s",
dbobj->dbus_path->str,
error->message);
} else {
NML_NMCLIENT_LOG_T(self,
"[%s] GetSettings() completed with success",
dbobj->dbus_path->str);
g_variant_get(ret, "(@a{sa{sv}})", &settings);
}
_nm_remote_settings_get_settings_commit(remote_connection, settings);
_dbus_handle_changes_commit(self, TRUE);
}
void
_nm_client_get_settings_call(NMClient *self, NMLDBusObject *dbobj)
{
GCancellable *cancellable;
cancellable = _nm_remote_settings_get_settings_prepare(NM_REMOTE_CONNECTION(dbobj->nmobj));
_nm_client_dbus_call_simple(self,
cancellable,
dbobj->dbus_path->str,
NM_DBUS_INTERFACE_SETTINGS_CONNECTION,
"GetSettings",
g_variant_new("()"),
G_VARIANT_TYPE("(a{sa{sv}})"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
_nm_client_get_settings_call_cb,
dbobj->nmobj);
}
static void
_dbus_settings_updated_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *object_path,
const char *signal_interface_name,
const char *signal_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *log_context = "settings-updated";
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->get_managed_objects_cancellable) {
/* we still wait for the initial GetManagedObjects(). Ignore the event. */
return;
}
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("()")))
return;
dbobj = _dbobjs_dbobj_get_s(self, object_path);
if (!dbobj || !NM_IS_REMOTE_CONNECTION(dbobj->nmobj)) {
NML_NMCLIENT_LOG_W(self,
"%s: [%s] ignore Updated signal for non-existing setting",
log_context,
object_path);
return;
}
NML_NMCLIENT_LOG_T(self, "%s: [%s] Updated signal received", log_context, object_path);
_nm_client_get_settings_call(self, dbobj);
}
/*****************************************************************************/
static void
_dbus_nm_connection_active_state_changed_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *object_path,
const char *signal_interface_name,
const char *signal_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *log_context = "active-connection-state-changed";
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
guint32 state;
guint32 reason;
if (priv->get_managed_objects_cancellable) {
/* we still wait for the initial GetManagedObjects(). Ignore the event. */
return;
}
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("(uu)"))) {
NML_NMCLIENT_LOG_E(self,
"%s: [%s] ignore StateChanged signal with unexpected signature",
log_context,
object_path);
return;
}
dbobj = _dbobjs_dbobj_get_s(self, object_path);
if (!dbobj || !NM_IS_ACTIVE_CONNECTION(dbobj->nmobj)) {
NML_NMCLIENT_LOG_E(self,
"%s: [%s] ignore StateChanged signal for non-existing active connection",
log_context,
object_path);
return;
}
g_variant_get(parameters, "(uu)", &state, &reason);
NML_NMCLIENT_LOG_T(self, "%s: [%s] StateChanged signal received", log_context, object_path);
_nm_active_connection_state_changed_commit(NM_ACTIVE_CONNECTION(dbobj->nmobj), state, reason);
_dbus_handle_changes_commit(self, TRUE);
}
/*****************************************************************************/
static void
_dbus_nm_vpn_connection_state_changed_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *object_path,
const char *signal_interface_name,
const char *signal_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *log_context = "vpn-connection-state-changed";
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
guint32 state;
guint32 reason;
if (priv->get_managed_objects_cancellable) {
/* we still wait for the initial GetManagedObjects(). Ignore the event. */
return;
}
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("(uu)"))) {
NML_NMCLIENT_LOG_E(self,
"%s: [%s] ignore VpnStateChanged signal with unexpected signature",
log_context,
object_path);
return;
}
dbobj = _dbobjs_dbobj_get_s(self, object_path);
if (!dbobj || !NM_IS_VPN_CONNECTION(dbobj->nmobj)) {
NML_NMCLIENT_LOG_E(self,
"%s: [%s] ignore VpnStateChanged signal for non-existing vpn connection",
log_context,
object_path);
return;
}
g_variant_get(parameters, "(uu)", &state, &reason);
NML_NMCLIENT_LOG_T(self, "%s: [%s] VpnStateChanged signal received", log_context, object_path);
_nm_vpn_connection_state_changed_commit(NM_VPN_CONNECTION(dbobj->nmobj), state, reason);
_dbus_handle_changes_commit(self, TRUE);
}
/*****************************************************************************/
static void
_emit_permissions_changed(NMClient *self, const guint8 *old_permissions, const guint8 *permissions)
{
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
int i;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (self->obj_base.is_disposing)
return;
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
if (old_permissions == permissions)
return;
for (i = 0; i < (int) G_N_ELEMENTS(nm_auth_permission_sorted); i++) {
NMClientPermission perm = nm_auth_permission_sorted[i];
NMClientPermissionResult perm_result = NM_CLIENT_PERMISSION_RESULT_UNKNOWN;
NMClientPermissionResult perm_result_old = NM_CLIENT_PERMISSION_RESULT_UNKNOWN;
if (permissions)
perm_result = permissions[perm - 1];
if (old_permissions)
perm_result_old = old_permissions[perm - 1];
if (perm_result == perm_result_old)
continue;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_signal_emit(self, signals[PERMISSION_CHANGED], 0, (guint) perm, (guint) perm_result);
}
}
static void
_dbus_check_permissions_start_cb(GObject *source, GAsyncResult *result, gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_auto_pop_gmaincontext GMainContext *dbus_context = NULL;
NMClient *self;
NMClientPrivate *priv;
gs_unref_variant GVariant *ret = NULL;
nm_auto_free_variant_iter GVariantIter *v_permissions = NULL;
gs_free guint8 *old_permissions = NULL;
gs_free_error GError *error = NULL;
const char *pkey;
const char *pvalue;
int i;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(source), result, &error);
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!ret && nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return;
self = user_data;
priv = NM_CLIENT_GET_PRIVATE(self);
g_clear_object(&priv->permissions_cancellable);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
old_permissions = g_steal_pointer(&priv->permissions);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!ret) {
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
/* when the call completes, we always pretend success. Even a failure means
* that we fetched the permissions, however they are all unknown. */
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(self, "GetPermissions call failed: %s", error->message);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
goto out;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
NML_NMCLIENT_LOG_T(self, "GetPermissions call finished with success");
g_variant_get(ret, "(a{ss})", &v_permissions);
while (g_variant_iter_next(v_permissions, "{&s&s}", &pkey, &pvalue)) {
NMClientPermission perm;
NMClientPermissionResult perm_result;
libnm,cli: cleanup mapping between NMClientPermission and strings
2019-12-09 08:02:01 +01:00
perm = nm_auth_permission_from_string(pkey);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (perm == NM_CLIENT_PERMISSION_NONE)
continue;
libnm,shared: move nm_permission_result_to_client() to shared's nm_client_permission_result_from_string()
2019-12-05 15:08:17 +01:00
perm_result = nm_client_permission_result_from_string(pvalue);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
if (!priv->permissions) {
if (perm_result == NM_CLIENT_PERMISSION_RESULT_UNKNOWN)
continue;
priv->permissions = g_new(guint8, G_N_ELEMENTS(nm_auth_permission_sorted));
for (i = 0; i < (int) G_N_ELEMENTS(nm_auth_permission_sorted); i++)
priv->permissions[i] = NM_CLIENT_PERMISSION_RESULT_UNKNOWN;
}
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
priv->permissions[perm - 1] = perm_result;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
out:
priv->permissions_state = NM_TERNARY_TRUE;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_context = nm_g_main_context_push_thread_default_if_necessary(priv->main_context);
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
_emit_permissions_changed(self, old_permissions, priv->permissions);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
_notify(self, PROP_PERMISSIONS_STATE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
static void
_dbus_check_permissions_start(NMClient *self)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
gboolean fetch;
fetch = !NM_FLAGS_HAS((NMClientInstanceFlags) priv->instance_flags,
NM_CLIENT_INSTANCE_FLAGS_NO_AUTO_FETCH_PERMISSIONS);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_cancellable(&priv->permissions_cancellable);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
if (fetch) {
NML_NMCLIENT_LOG_T(self, "GetPermissions() call started...");
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
priv->permissions_cancellable = g_cancellable_new();
_nm_client_dbus_call_simple(self,
priv->permissions_cancellable,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"GetPermissions",
g_variant_new("()"),
G_VARIANT_TYPE("(a{ss})"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
_dbus_check_permissions_start_cb,
self);
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
static void
_dbus_nm_check_permissions_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *object_path,
const char *signal_interface_name,
const char *signal_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("()"))) {
NML_NMCLIENT_LOG_E(self,
"ignore CheckPermissions signal with unexpected signature %s",
g_variant_get_type_string(parameters));
return;
}
_dbus_check_permissions_start(self);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
if (priv->permissions_state == NM_TERNARY_TRUE)
priv->permissions_state = NM_TERNARY_FALSE;
_notify(self, PROP_PERMISSIONS_STATE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
/*****************************************************************************/
static void
_property_ao_notify_changed_connections_cb(NMLDBusPropertyAO *pr_ao,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self,
NMObject *nmobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean is_added /* or else removed */)
{
_nm_client_notify_event_queue_emit_obj_signal(self,
G_OBJECT(self),
nmobj,
is_added,
5,
is_added ? signals[CONNECTION_ADDED]
: signals[CONNECTION_REMOVED]);
}
static void
_property_ao_notify_changed_all_devices_cb(NMLDBusPropertyAO *pr_ao,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self,
NMObject *nmobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean is_added /* or else removed */)
{
_nm_client_notify_event_queue_emit_obj_signal(self,
G_OBJECT(self),
nmobj,
is_added,
6,
is_added ? signals[ANY_DEVICE_ADDED]
: signals[ANY_DEVICE_REMOVED]);
}
static void
_property_ao_notify_changed_devices_cb(NMLDBusPropertyAO *pr_ao,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self,
NMObject *nmobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean is_added /* or else removed */)
{
_nm_client_notify_event_queue_emit_obj_signal(self,
G_OBJECT(self),
nmobj,
is_added,
7,
is_added ? signals[DEVICE_ADDED]
: signals[DEVICE_REMOVED]);
}
static void
_property_ao_notify_changed_active_connections_cb(NMLDBusPropertyAO *pr_ao,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self,
NMObject *nmobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean is_added /* or else removed */)
{
_nm_client_notify_event_queue_emit_obj_signal(self,
G_OBJECT(self),
nmobj,
is_added,
8,
is_added ? signals[ACTIVE_CONNECTION_ADDED]
: signals[ACTIVE_CONNECTION_REMOVED]);
}
/*****************************************************************************/
typedef struct {
NMLDBusObjWatcherWithPtr *obj_watcher;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *op_name;
NMLDBusObject *dbobj;
GTask *task;
GVariant *extra_results;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer result;
GType gtype;
gulong cancellable_id;
} RequestWaitData;
static void
_request_wait_data_free(RequestWaitData *request_data)
{
nm_assert(!request_data->obj_watcher);
nm_assert(request_data->cancellable_id == 0);
nm_assert(!request_data->task || G_IS_TASK(request_data->task));
nm_g_object_unref(request_data->task);
nm_g_object_unref(request_data->result);
nm_g_variant_unref(request_data->extra_results);
if (request_data->dbobj)
nml_dbus_object_unref(request_data->dbobj);
nm_g_slice_free(request_data);
}
static void
_request_wait_task_return(RequestWaitData *request_data)
{
gs_unref_object GTask *task = NULL;
nm_assert(request_data);
nm_assert(G_IS_TASK(request_data->task));
nm_assert(request_data->dbobj);
nm_assert(NM_IS_OBJECT(request_data->dbobj->nmobj));
nm_assert(!request_data->result);
task = g_steal_pointer(&request_data->task);
request_data->result = g_object_ref(request_data->dbobj->nmobj);
nm_clear_g_signal_handler(g_task_get_cancellable(task), &request_data->cancellable_id);
nm_clear_pointer(&request_data->dbobj, nml_dbus_object_unref);
g_task_return_pointer(task, request_data, (GDestroyNotify) _request_wait_data_free);
}
static gboolean
_request_wait_complete(NMClient *self, RequestWaitData *request_data, gboolean force_complete)
{
NMLDBusObject *dbobj;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(request_data);
nm_assert(!request_data->result);
nm_assert(!request_data->obj_watcher);
nm_assert(request_data->dbobj);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbobj = request_data->dbobj;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (dbobj->obj_state == NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY) {
NML_NMCLIENT_LOG_D(self,
"%s() succeeded with %s",
request_data->op_name,
dbobj->dbus_path->str);
nm_assert(G_TYPE_CHECK_INSTANCE_TYPE(dbobj->nmobj, request_data->gtype));
_request_wait_task_return(request_data);
return TRUE;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (force_complete || dbobj->obj_state != NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY) {
if (force_complete)
NML_NMCLIENT_LOG_D(self,
"%s() succeeded with %s but object is in an unsuitable state",
request_data->op_name,
dbobj->dbus_path->str);
else
NML_NMCLIENT_LOG_W(self,
"%s() succeeded with %s but object is in an unsuitable state",
request_data->op_name,
dbobj->dbus_path->str);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_task_return_error(
request_data->task,
g_error_new(NM_CLIENT_ERROR,
NM_CLIENT_ERROR_OBJECT_CREATION_FAILED,
_("request succeeded with %s but object is in an unsuitable state"),
dbobj->dbus_path->str));
_request_wait_data_free(request_data);
return TRUE;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return FALSE;
}
static void
_request_wait_complete_cb(NMClient *self, NMClientNotifyEventWithPtr *notify_event)
{
_request_wait_complete(self, notify_event->user_data, TRUE);
}
static void
_request_wait_obj_watcher_cb(NMClient *self, gpointer obj_watcher_base)
{
NMLDBusObjWatcherWithPtr *obj_watcher = obj_watcher_base;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
RequestWaitData *request_data = obj_watcher->user_data;
NMLDBusObject *dbobj;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbobj = request_data->dbobj;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (dbobj->obj_state == NML_DBUS_OBJ_STATE_WITH_NMOBJ_NOT_READY)
return;
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IN_SET((NMLDBusObjState) dbobj->obj_state,
NML_DBUS_OBJ_STATE_WATCHED_ONLY,
NML_DBUS_OBJ_STATE_ON_DBUS,
NML_DBUS_OBJ_STATE_WITH_NMOBJ_READY));
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_dbobjs_obj_watcher_unregister(self, g_steal_pointer(&request_data->obj_watcher));
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_signal_handler(g_task_get_cancellable(request_data->task),
&request_data->cancellable_id);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_notify_event_queue_with_ptr(self,
NM_CLIENT_NOTIFY_EVENT_PRIO_AFTER + 30,
_request_wait_complete_cb,
request_data);
}
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static void
_request_wait_cancelled_cb(GCancellable *cancellable, gpointer user_data)
{
RequestWaitData *request_data = user_data;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self;
GError *error = NULL;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(cancellable == g_task_get_cancellable(request_data->task));
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_utils_error_set_cancelled(&error, FALSE, NULL);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
self = g_task_get_source_object(request_data->task);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_signal_handler(cancellable, &request_data->cancellable_id);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_dbobjs_obj_watcher_unregister(self, g_steal_pointer(&request_data->obj_watcher));
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_task_return_error(request_data->task, error);
libnm: add helper functions for refactoring D-Bus calls in libnm We will drop GDBusProxy and the gdbus-codegen classes. First, we need to replace all D-Bus calls from nmdbus_*() API with plain uses of GDBusConnection. For that, add accessors to get the dbus-connection and the name-owner. This API is not beautiful, it's an interim solution for now.
2019-10-04 13:56:36 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_request_wait_data_free(request_data);
libnm: add helper functions for refactoring D-Bus calls in libnm We will drop GDBusProxy and the gdbus-codegen classes. First, we need to replace all D-Bus calls from nmdbus_*() API with plain uses of GDBusConnection. For that, add accessors to get the dbus-connection and the name-owner. This API is not beautiful, it's an interim solution for now.
2019-10-04 13:56:36 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_request_wait_start(GTask *task_take,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const char *op_name,
GType gtype,
const char *dbus_path,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *extra_results_take)
libnm: add helper functions for refactoring D-Bus calls in libnm We will drop GDBusProxy and the gdbus-codegen classes. First, we need to replace all D-Bus calls from nmdbus_*() API with plain uses of GDBusConnection. For that, add accessors to get the dbus-connection and the name-owner. This API is not beautiful, it's an interim solution for now.
2019-10-04 13:56:36 +02:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_object GTask *task = g_steal_pointer(&task_take);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
RequestWaitData *request_data;
GCancellable *cancellable;
NMLDBusObject *dbobj;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(G_IS_TASK(task));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
self = g_task_get_source_object(task);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbobj = _dbobjs_get_nmobj(self, dbus_path, gtype);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!dbobj) {
NML_NMCLIENT_LOG_E(self,
"%s() succeeded with %s but object does not exist",
op_name,
dbus_path);
g_task_return_error(task,
g_error_new(NM_CLIENT_ERROR,
NM_CLIENT_ERROR_FAILED,
_("operation succeeded but object %s does not exist"),
dbus_path));
return;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
request_data = g_slice_new(RequestWaitData);
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
*request_data = (RequestWaitData) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
.task = g_steal_pointer(&task),
.op_name = op_name,
.gtype = gtype,
.dbobj = nml_dbus_object_ref(dbobj),
.obj_watcher = NULL,
.extra_results = g_steal_pointer(&extra_results_take),
.result = NULL,
.cancellable_id = 0,
};
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (_request_wait_complete(self, request_data, FALSE))
return;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(self,
"%s() succeeded with %s. Wait for object to become ready",
op_name,
dbobj->dbus_path->str);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
request_data->obj_watcher = _dbobjs_obj_watcher_register_o(self,
dbobj,
_request_wait_obj_watcher_cb,
sizeof(NMLDBusObjWatcherWithPtr));
request_data->obj_watcher->user_data = request_data;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
cancellable = g_task_get_cancellable(request_data->task);
if (cancellable) {
gulong id;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
id = g_cancellable_connect(cancellable,
G_CALLBACK(_request_wait_cancelled_cb),
request_data,
NULL);
if (id == 0) {
/* the callback was invoked synchronously, which destroyed @request_data.
* We must not touch @info anymore. */
} else
request_data->cancellable_id = id;
}
}
libnm: add helper functions for refactoring D-Bus calls in libnm We will drop GDBusProxy and the gdbus-codegen classes. First, we need to replace all D-Bus calls from nmdbus_*() API with plain uses of GDBusConnection. For that, add accessors to get the dbus-connection and the name-owner. This API is not beautiful, it's an interim solution for now.
2019-10-04 13:56:36 +02:00
libnm: add assertion for object returned by nm_device_get_active_connection() I have a coredump that seems to indicate that nm_device_get_active_connection() did not return a valid object. Let's add an assertion, trying to identify the issue earlier. Aside from that, this change isn't useful, but an nm_assert() shouldn't hurt anyway.
2019-11-28 12:40:55 +01:00
static gpointer
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_request_wait_finish(NMClient *client,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GAsyncResult *result,
gpointer source_tag,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant **out_result,
GError **error)
libnm: clean up NMClient semantics when NM is not running Add a single function to check if NM is running and set a GError if not, then use it as appropriate. Don't bother to check if NM is running in get_*() functions if nm_running_changed_cb() would have reset the field anyway (and fix that up to reset a few more fields).
2014-10-07 09:26:40 -04:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
RequestWaitData *request_data = NULL;
gpointer r;
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
g_return_val_if_fail(nm_g_task_is_valid(result, client, source_tag), NULL);
request_data = g_task_propagate_pointer(G_TASK(result), error);
if (!request_data) {
NM_SET_OUT(out_result, NULL);
return NULL;
libnm: clean up NMClient semantics when NM is not running Add a single function to check if NM is running and set a GError if not, then use it as appropriate. Don't bother to check if NM is running in get_*() functions if nm_running_changed_cb() would have reset the field anyway (and fix that up to reset a few more fields).
2014-10-07 09:26:40 -04:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_OBJECT(request_data->result));
NM_SET_OUT(out_result, g_steal_pointer(&request_data->extra_results));
r = g_steal_pointer(&request_data->result);
libnm: add assertion for object returned by nm_device_get_active_connection() I have a coredump that seems to indicate that nm_device_get_active_connection() did not return a valid object. Let's add an assertion, trying to identify the issue earlier. Aside from that, this change isn't useful, but an nm_assert() shouldn't hurt anyway.
2019-11-28 12:40:55 +01:00
nm_assert(NM_IS_OBJECT(r));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_request_wait_data_free(request_data);
return r;
libnm: clean up NMClient semantics when NM is not running Add a single function to check if NM is running and set a GError if not, then use it as appropriate. Don't bother to check if NM is running in get_*() functions if nm_running_changed_cb() would have reset the field anyway (and fix that up to reset a few more fields).
2014-10-07 09:26:40 -04:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
/**
* nm_client_get_instance_flags:
* @self: the #NMClient instance.
*
* Returns: the #NMClientInstanceFlags flags.
*
* Since: 1.24
*/
NMClientInstanceFlags
nm_client_get_instance_flags(NMClient *self)
{
g_return_val_if_fail(NM_IS_CLIENT(self), NM_CLIENT_INSTANCE_FLAGS_NONE);
return NM_CLIENT_GET_PRIVATE(self)->instance_flags;
}
libnm: add NM_CLIENT_DBUS_CONNECTION property The used GDBusConnection should be configurable when creating the NMClient instance. Automatically choosing one of the g_bus_get() singletons is fine by default, but it's an unnecessary limitation.
2019-10-12 14:59:23 +02:00
/**
* nm_client_get_dbus_connection:
* @client: a #NMClient
*
* Gets the %GDBusConnection of the instance. This can be either passed when
* constructing the instance (as "dbus-connection" property), or it will be
* automatically initialized during async/sync init.
*
* Returns: (transfer none): the D-Bus connection of the client, or %NULL if none is set.
*
* Since: 1.22
**/
GDBusConnection *
nm_client_get_dbus_connection(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
return NM_CLIENT_GET_PRIVATE(client)->dbus_connection;
}
libnm: add NM_CLIENT_DBUS_NAME_OWNER property It's not yet implemented. But obviously it's interesting to get the name owner to which the NMClient is currently connected. Note only that: the name-owner property really says whether NM is currently running or not.
2019-10-15 16:08:36 +02:00
/**
* nm_client_get_dbus_name_owner:
* @client: a #NMClient
*
* Returns: (transfer none): the current name owner of the D-Bus service of NetworkManager.
*
* Since: 1.22
**/
const char *
nm_client_get_dbus_name_owner(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->name_owner;
libnm: add NM_CLIENT_DBUS_NAME_OWNER property It's not yet implemented. But obviously it's interesting to get the name owner to which the NMClient is currently connected. Note only that: the name-owner property really says whether NM is currently running or not.
2019-10-15 16:08:36 +02:00
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_version:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets NetworkManager version.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: clean up NMClient semantics when NM is not running Add a single function to check if NM is running and set a GError if not, then use it as appropriate. Don't bother to check if NM is running in get_*() functions if nm_running_changed_cb() would have reset the field anyway (and fix that up to reset a few more fields).
2014-10-07 09:26:40 -04:00
* Returns: string with the version (or %NULL if NetworkManager is not running)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
const char *
nm_client_get_version(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.version;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_state:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets the current daemon state.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: the current %NMState
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMState
nm_client_get_state(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NM_STATE_UNKNOWN);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.state;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_startup:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Tests whether the daemon is still in the process of activating
* connections at startup.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: whether the daemon is still starting up
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
gboolean
nm_client_get_startup(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.startup;
}
static void
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
_set_nm_running(NMClient *self, gboolean queue_notify)
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
gboolean nm_running;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
nm_running = priv->dbobj_nm && priv->name_owner && !priv->get_managed_objects_cancellable;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->nm_running != nm_running) {
priv->nm_running = nm_running;
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
if (queue_notify) {
_nm_client_queue_notify_object(self, self, obj_properties[PROP_NM_RUNNING]);
} else
_notify(self, PROP_NM_RUNNING);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
/**
* nm_client_get_nm_running:
* @client: a #NMClient
*
* Determines whether the daemon is running.
*
* Returns: %TRUE if the daemon is running
**/
gboolean
nm_client_get_nm_running(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm_running;
}
client: add nm_client_get_object_by_path() and nm_object_get_client() API When iterating the GMainContext of the NMClient instance, D-Bus events get processed. That means, every time you iterate the context (or "return to the main loop"), the content of the cache might change completely. It makes sense to keep a reference to an NMObject instance, do something, and afterwards check whether the instance can still be found in the cache. Add an API for that. nm_object_get_client() allows to know whether the object is still cached. Likewise, while NMClient abstracts D-Bus, it should still provide a way to look up an NMObject by D-Bus path. Add nm_client_get_object_by_path() for that. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/384
2020-01-04 07:58:58 +01:00
/**
* nm_client_get_object_by_path:
* @client: the #NMClient instance
* @dbus_path: the D-Bus path of the object to look up
*
* Returns: (transfer none): the #NMObject instance that is
* cached under @dbus_path, or %NULL if no such object exists.
*
* Since: 1.24
*/
NMObject *
nm_client_get_object_by_path(NMClient *client, const char *dbus_path)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
g_return_val_if_fail(dbus_path, NULL);
return _dbobjs_get_nmobj_unpack_visible(client, dbus_path, G_TYPE_NONE);
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/**
* nm_client_get_metered:
* @client: a #NMClient
*
* Returns: whether the default route is metered.
*
* Since: 1.22
*/
NMMetered
nm_client_get_metered(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NM_METERED_UNKNOWN);
return NM_CLIENT_GET_PRIVATE(client)->nm.metered;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
/**
* nm_client_networking_get_enabled:
* @client: a #NMClient
*
* Whether networking is enabled or disabled.
*
* Returns: %TRUE if networking is enabled, %FALSE if networking is disabled
**/
gboolean
nm_client_networking_get_enabled(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.networking_enabled;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
/**
* nm_client_networking_set_enabled:
* @client: a #NMClient
* @enabled: %TRUE to set networking enabled, %FALSE to set networking disabled
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @error: return location for a #GError, or %NULL
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
*
* Enables or disables networking. When networking is disabled, all controlled
* interfaces are disconnected and deactivated. When networking is enabled,
* all controlled interfaces are available for activation.
*
* Returns: %TRUE on success, %FALSE otherwise
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm: advise using D-Bus instead of deprecated synchronous methods With 1.22, various synchronous functions for invoking D-Bus methods were deprecated. The reason was that D-Bus is fundamentally asynchronous, and providing synchronous API in NMClient is inherently wrong. That is because NMClient essentially is a cache of the D-Bus API, and invoking g_dbus_connection_call_sync() messes up the order of events from D-Bus. In particular, when the synchronous function completes, the content of the cache does not yet reflect the change. Since they got deprecated, the question is with what to replace them. Instead of adding a (e.g.) nm_client_networking_set_enabled_async() for nm_client_networking_set_enabled(), just expect the user to call D-Bus directly. D-Bus itself defines a reasonable API, and with GDBusConnection it is fine (and convenient) to just call D-Bus operations directly. Often libraries try to abstract D-Bus by providing convenience wrappers around D-Bus API. I think that often is wrong and unnecessary. Note that libnm's NMClient does a lot more than just wrapping simple D-Bus calls. It provides a complete client-side cache of the D-Bus interface. As such, what libnm's NMClient does is more than simple wrappers around D-Bus. NMClient is a reasonable thing to do. However, it is unnecessary to add API like nm_client_networking_set_enabled_async() that only calls g_dbus_connection_call(). Don't pretend that we would need such trivial wrappers in libnm. Instead, recommend to use g_dbus_connection_call(). Or alternatively, the convenience wrappers nm_client_dbus_call() and nm_client_dbus_set_property().
2020-03-13 18:24:37 +01:00
* Deprecated: 1.22: Use the async command nm_client_dbus_call() on %NM_DBUS_PATH,
* %NM_DBUS_INTERFACE to call "Enable" with "(b)" arguments and no return value.
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
**/
gboolean
nm_client_networking_set_enabled(NMClient *client, gboolean enable, GError **error)
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return _nm_client_dbus_call_sync_void(client,
NULL,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"Enable",
g_variant_new("(b)", enable),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
}
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
/**
* nm_client_wireless_get_enabled:
* @client: a #NMClient
*
* Determines whether the wireless is enabled.
*
* Returns: %TRUE if wireless is enabled
**/
gboolean
nm_client_wireless_get_enabled(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.wireless_enabled;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_wireless_set_enabled:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @enabled: %TRUE to enable wireless
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Enables or disables wireless devices.
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm: advise using D-Bus instead of deprecated synchronous methods With 1.22, various synchronous functions for invoking D-Bus methods were deprecated. The reason was that D-Bus is fundamentally asynchronous, and providing synchronous API in NMClient is inherently wrong. That is because NMClient essentially is a cache of the D-Bus API, and invoking g_dbus_connection_call_sync() messes up the order of events from D-Bus. In particular, when the synchronous function completes, the content of the cache does not yet reflect the change. Since they got deprecated, the question is with what to replace them. Instead of adding a (e.g.) nm_client_networking_set_enabled_async() for nm_client_networking_set_enabled(), just expect the user to call D-Bus directly. D-Bus itself defines a reasonable API, and with GDBusConnection it is fine (and convenient) to just call D-Bus operations directly. Often libraries try to abstract D-Bus by providing convenience wrappers around D-Bus API. I think that often is wrong and unnecessary. Note that libnm's NMClient does a lot more than just wrapping simple D-Bus calls. It provides a complete client-side cache of the D-Bus interface. As such, what libnm's NMClient does is more than simple wrappers around D-Bus. NMClient is a reasonable thing to do. However, it is unnecessary to add API like nm_client_networking_set_enabled_async() that only calls g_dbus_connection_call(). Don't pretend that we would need such trivial wrappers in libnm. Instead, recommend to use g_dbus_connection_call(). Or alternatively, the convenience wrappers nm_client_dbus_call() and nm_client_dbus_set_property().
2020-03-13 18:24:37 +01:00
* Deprecated: 1.22: Use the async command nm_client_dbus_set_property() on %NM_DBUS_PATH,
* %NM_DBUS_INTERFACE to set "WirelessEnabled" property to a "(b)" value.
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*/
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
void
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
nm_client_wireless_set_enabled(NMClient *client, gboolean enabled)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_set_property_sync_legacy(client,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"WirelessEnabled",
"b",
enabled);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_wireless_hardware_get_enabled:
* @client: a #NMClient
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Determines whether the wireless hardware is enabled.
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: %TRUE if the wireless hardware is enabled
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
gboolean
nm_client_wireless_hardware_get_enabled(NMClient *client)
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.wireless_hardware_enabled;
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
}
libnm: support radio flags
2022-03-21 10:20:11 +01:00
/**
* nm_client_get_radio_flags:
* @client: a #NMClient
*
* Get radio flags.
*
* Returns: the #NMRadioFlags.
*
* Since: 1.38
**/
NMRadioFlags
nm_client_get_radio_flags(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NM_RADIO_FLAG_NONE);
return NM_CLIENT_GET_PRIVATE(client)->nm.radio_flags;
}
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
/**
* nm_client_wwan_get_enabled:
* @client: a #NMClient
*
* Determines whether WWAN is enabled.
*
* Returns: %TRUE if WWAN is enabled
**/
gboolean
nm_client_wwan_get_enabled(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.wwan_enabled;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_wwan_set_enabled:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @enabled: %TRUE to enable WWAN
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Enables or disables WWAN devices.
libnm: advise using D-Bus instead of deprecated synchronous methods With 1.22, various synchronous functions for invoking D-Bus methods were deprecated. The reason was that D-Bus is fundamentally asynchronous, and providing synchronous API in NMClient is inherently wrong. That is because NMClient essentially is a cache of the D-Bus API, and invoking g_dbus_connection_call_sync() messes up the order of events from D-Bus. In particular, when the synchronous function completes, the content of the cache does not yet reflect the change. Since they got deprecated, the question is with what to replace them. Instead of adding a (e.g.) nm_client_networking_set_enabled_async() for nm_client_networking_set_enabled(), just expect the user to call D-Bus directly. D-Bus itself defines a reasonable API, and with GDBusConnection it is fine (and convenient) to just call D-Bus operations directly. Often libraries try to abstract D-Bus by providing convenience wrappers around D-Bus API. I think that often is wrong and unnecessary. Note that libnm's NMClient does a lot more than just wrapping simple D-Bus calls. It provides a complete client-side cache of the D-Bus interface. As such, what libnm's NMClient does is more than simple wrappers around D-Bus. NMClient is a reasonable thing to do. However, it is unnecessary to add API like nm_client_networking_set_enabled_async() that only calls g_dbus_connection_call(). Don't pretend that we would need such trivial wrappers in libnm. Instead, recommend to use g_dbus_connection_call(). Or alternatively, the convenience wrappers nm_client_dbus_call() and nm_client_dbus_set_property().
2020-03-13 18:24:37 +01:00
*
* Deprecated: 1.22: Use the async command nm_client_dbus_set_property() on %NM_DBUS_PATH,
* %NM_DBUS_INTERFACE to set "WwanEnabled" property to a "(b)" value.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
void
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
nm_client_wwan_set_enabled(NMClient *client, gboolean enabled)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_set_property_sync_legacy(client,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"WwanEnabled",
"b",
enabled);
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_wwan_hardware_get_enabled:
* @client: a #NMClient
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Determines whether the WWAN hardware is enabled.
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: %TRUE if the WWAN hardware is enabled
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
gboolean
nm_client_wwan_hardware_get_enabled(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.wwan_hardware_enabled;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_wimax_get_enabled:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Determines whether WiMAX is enabled.
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: %TRUE if WiMAX is enabled
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: This function always returns FALSE because WiMax is no longer supported.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
gboolean
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
nm_client_wimax_get_enabled(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: make sync/async APIs more GLib-like Make synchronous APIs take GCancellables, and make asynchronous APIs use GAsyncReadyCallbacks and have names ending in "_async", with "_finish" functions to retrieve the results. Also, make nm_client_activate_connection_finish(), nm_client_add_and_activate_finish(), and nm_remote_settings_add_connection_finish() be (transfer full) rather than (transfer none), because the refcounting semantics become slightly confusing in some edge cases otherwise.
2014-09-11 16:27:13 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
return FALSE;
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_wimax_set_enabled:
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @enabled: %TRUE to enable WiMAX
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Enables or disables WiMAX devices.
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: This function does nothing because WiMax is no longer supported.
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
**/
void
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
nm_client_wimax_set_enabled(NMClient *client, gboolean enabled)
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
{
g_return_if_fail(NM_IS_CLIENT(client));
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_wimax_hardware_get_enabled:
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
* @client: a #NMClient
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Determines whether the WiMAX hardware is enabled.
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: %TRUE if the WiMAX hardware is enabled
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: This function always returns FALSE because WiMax is no longer supported.
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
**/
gboolean
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
nm_client_wimax_hardware_get_enabled(NMClient *client)
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
return FALSE;
libnm: add some missing sync/async method variants Add the missing variant in most places in the API where previously there was either only a synchronous version or only an asynchronous version. There is not yet a synchronous nm_client_activate_connection(), nm_client_add_and_activate_connection(), or nm_remote_settings_add_connection(), because the existing async code depends on waiting for other asynchronous events, so making them run synchronously is slightly more complicated. But these APIs can be added later.
2014-09-11 16:27:13 -04:00
}
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
/**
* nm_client_connectivity_check_get_available:
* @client: a #NMClient
*
* Determine whether connectivity checking is available. This
* requires that the URI of a connectivity service has been set in the
* configuration file.
*
* Returns: %TRUE if connectivity checking is available.
libnm: add gtk-doc Since markers for new API and missing enum documentation
2017-08-17 23:08:44 +02:00
*
* Since: 1.10
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
*/
gboolean
nm_client_connectivity_check_get_available(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.connectivity_check_available;
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
}
/**
* nm_client_connectivity_check_get_enabled:
* @client: a #NMClient
*
* Determine whether connectivity checking is enabled.
*
* Returns: %TRUE if connectivity checking is enabled.
libnm: add gtk-doc Since markers for new API and missing enum documentation
2017-08-17 23:08:44 +02:00
*
* Since: 1.10
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
*/
gboolean
nm_client_connectivity_check_get_enabled(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.connectivity_check_enabled;
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
}
/**
* nm_client_connectivity_check_set_enabled:
* @client: a #NMClient
* @enabled: %TRUE to enable connectivity checking
*
* Enable or disable connectivity checking. Note that if a
* connectivity checking URI has not been configured, this will not
* have any effect.
libnm: add gtk-doc Since markers for new API and missing enum documentation
2017-08-17 23:08:44 +02:00
*
* Since: 1.10
libnm: advise using D-Bus instead of deprecated synchronous methods With 1.22, various synchronous functions for invoking D-Bus methods were deprecated. The reason was that D-Bus is fundamentally asynchronous, and providing synchronous API in NMClient is inherently wrong. That is because NMClient essentially is a cache of the D-Bus API, and invoking g_dbus_connection_call_sync() messes up the order of events from D-Bus. In particular, when the synchronous function completes, the content of the cache does not yet reflect the change. Since they got deprecated, the question is with what to replace them. Instead of adding a (e.g.) nm_client_networking_set_enabled_async() for nm_client_networking_set_enabled(), just expect the user to call D-Bus directly. D-Bus itself defines a reasonable API, and with GDBusConnection it is fine (and convenient) to just call D-Bus operations directly. Often libraries try to abstract D-Bus by providing convenience wrappers around D-Bus API. I think that often is wrong and unnecessary. Note that libnm's NMClient does a lot more than just wrapping simple D-Bus calls. It provides a complete client-side cache of the D-Bus interface. As such, what libnm's NMClient does is more than simple wrappers around D-Bus. NMClient is a reasonable thing to do. However, it is unnecessary to add API like nm_client_networking_set_enabled_async() that only calls g_dbus_connection_call(). Don't pretend that we would need such trivial wrappers in libnm. Instead, recommend to use g_dbus_connection_call(). Or alternatively, the convenience wrappers nm_client_dbus_call() and nm_client_dbus_set_property().
2020-03-13 18:24:37 +01:00
*
* Deprecated: 1.22: Use the async command nm_client_dbus_set_property() on %NM_DBUS_PATH,
* %NM_DBUS_INTERFACE to set "ConnectivityCheckEnabled" property to a "(b)" value.
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
*/
void
nm_client_connectivity_check_set_enabled(NMClient *client, gboolean enabled)
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_set_property_sync_legacy(client,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"ConnectivityCheckEnabled",
"b",
enabled);
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
}
libnm,core: Add ConnectivityCheckUri property and accessors So that applications like GNOME Shell can hit the same URI to show the captive portal login page. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/209
2019-07-22 15:55:15 +01:00
/**
* nm_client_connectivity_check_get_uri:
* @client: a #NMClient
*
* Get the URI that will be queried to determine if there is internet
* connectivity.
*
* Returns: (transfer none): the connectivity URI in use
*
* Since: 1.20
*/
const char *
nm_client_connectivity_check_get_uri(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.connectivity_check_uri;
libnm,core: Add ConnectivityCheckUri property and accessors So that applications like GNOME Shell can hit the same URI to show the captive portal login page. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/209
2019-07-22 15:55:15 +01:00
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_logging:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
doc: reorder gtkdoc annotations for consistency
2023-03-02 12:18:30 +01:00
* @level: (out) (optional) (nullable): return location for logging level string
* @domains: (out) (optional) (nullable): return location for log domains string. The string is
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* a list of domains separated by ","
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @error: return location for a #GError, or %NULL
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets NetworkManager current logging level and domains.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: %TRUE on success, %FALSE otherwise
libnm: mark more synchronous libnm API as deprecated This is a follow-up to commit e90684a169a7 ('libnm: deprecate synchronous/blocking API in libnm') to mark more of such synchronous API as deprecated.
2019-10-04 16:48:39 +02:00
*
libnm: advise using D-Bus instead of deprecated synchronous methods With 1.22, various synchronous functions for invoking D-Bus methods were deprecated. The reason was that D-Bus is fundamentally asynchronous, and providing synchronous API in NMClient is inherently wrong. That is because NMClient essentially is a cache of the D-Bus API, and invoking g_dbus_connection_call_sync() messes up the order of events from D-Bus. In particular, when the synchronous function completes, the content of the cache does not yet reflect the change. Since they got deprecated, the question is with what to replace them. Instead of adding a (e.g.) nm_client_networking_set_enabled_async() for nm_client_networking_set_enabled(), just expect the user to call D-Bus directly. D-Bus itself defines a reasonable API, and with GDBusConnection it is fine (and convenient) to just call D-Bus operations directly. Often libraries try to abstract D-Bus by providing convenience wrappers around D-Bus API. I think that often is wrong and unnecessary. Note that libnm's NMClient does a lot more than just wrapping simple D-Bus calls. It provides a complete client-side cache of the D-Bus interface. As such, what libnm's NMClient does is more than simple wrappers around D-Bus. NMClient is a reasonable thing to do. However, it is unnecessary to add API like nm_client_networking_set_enabled_async() that only calls g_dbus_connection_call(). Don't pretend that we would need such trivial wrappers in libnm. Instead, recommend to use g_dbus_connection_call(). Or alternatively, the convenience wrappers nm_client_dbus_call() and nm_client_dbus_set_property().
2020-03-13 18:24:37 +01:00
* Deprecated: 1.22: Use the async command nm_client_dbus_call() on %NM_DBUS_PATH,
* %NM_DBUS_INTERFACE to call "GetLogging" with no arguments to get "(ss)" for level
* and domains.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
gboolean
libnm: implement nm_client_get_logging()/nm_client_set_logging() by using GDBusConnection directly
2019-10-04 16:18:51 +02:00
nm_client_get_logging(NMClient *client, char **level, char **domains, GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: implement nm_client_get_logging()/nm_client_set_logging() by using GDBusConnection directly
2019-10-04 16:18:51 +02:00
gs_unref_variant GVariant *ret = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
g_return_val_if_fail(level == NULL || *level == NULL, FALSE);
g_return_val_if_fail(domains == NULL || *domains == NULL, FALSE);
g_return_val_if_fail(error == NULL || *error == NULL, FALSE);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = _nm_client_dbus_call_sync(client,
libnm: implement nm_client_get_logging()/nm_client_set_logging() by using GDBusConnection directly
2019-10-04 16:18:51 +02:00
NULL,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"GetLogging",
g_variant_new("()"),
G_VARIANT_TYPE("(ss)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
if (!ret)
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
return FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_get_logging()/nm_client_set_logging() by using GDBusConnection directly
2019-10-04 16:18:51 +02:00
g_variant_get(ret, "(ss)", level, domains);
return TRUE;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_set_logging:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @level: (nullable): logging level to set (%NULL or an empty string for no change)
* @domains: (nullable): logging domains to set. The string should be a list of log
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* domains separated by ",". (%NULL or an empty string for no change)
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @error: return location for a #GError, or %NULL
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Sets NetworkManager logging level and/or domains.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: %TRUE on success, %FALSE otherwise
libnm: mark more synchronous libnm API as deprecated This is a follow-up to commit e90684a169a7 ('libnm: deprecate synchronous/blocking API in libnm') to mark more of such synchronous API as deprecated.
2019-10-04 16:48:39 +02:00
*
libnm: advise using D-Bus instead of deprecated synchronous methods With 1.22, various synchronous functions for invoking D-Bus methods were deprecated. The reason was that D-Bus is fundamentally asynchronous, and providing synchronous API in NMClient is inherently wrong. That is because NMClient essentially is a cache of the D-Bus API, and invoking g_dbus_connection_call_sync() messes up the order of events from D-Bus. In particular, when the synchronous function completes, the content of the cache does not yet reflect the change. Since they got deprecated, the question is with what to replace them. Instead of adding a (e.g.) nm_client_networking_set_enabled_async() for nm_client_networking_set_enabled(), just expect the user to call D-Bus directly. D-Bus itself defines a reasonable API, and with GDBusConnection it is fine (and convenient) to just call D-Bus operations directly. Often libraries try to abstract D-Bus by providing convenience wrappers around D-Bus API. I think that often is wrong and unnecessary. Note that libnm's NMClient does a lot more than just wrapping simple D-Bus calls. It provides a complete client-side cache of the D-Bus interface. As such, what libnm's NMClient does is more than simple wrappers around D-Bus. NMClient is a reasonable thing to do. However, it is unnecessary to add API like nm_client_networking_set_enabled_async() that only calls g_dbus_connection_call(). Don't pretend that we would need such trivial wrappers in libnm. Instead, recommend to use g_dbus_connection_call(). Or alternatively, the convenience wrappers nm_client_dbus_call() and nm_client_dbus_set_property().
2020-03-13 18:24:37 +01:00
* Deprecated: 1.22: Use the async command nm_client_dbus_call() on %NM_DBUS_PATH,
* %NM_DBUS_INTERFACE to call "SetLogging" with "(ss)" arguments for level and domains.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
gboolean
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
nm_client_set_logging(NMClient *client, const char *level, const char *domains, GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(error == NULL || *error == NULL, FALSE);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return _nm_client_dbus_call_sync_void(client,
libnm: implement nm_client_get_logging()/nm_client_set_logging() by using GDBusConnection directly
2019-10-04 16:18:51 +02:00
NULL,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"SetLogging",
g_variant_new("(ss)", level ?: "", domains ?: ""),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_permission_result:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @permission: the permission for which to return the result, one of #NMClientPermission
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Requests the result of a specific permission, which indicates whether the
* client can or cannot perform the action the permission represents
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: the permission's result, one of #NMClientPermissionResult
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMClientPermissionResult
nm_client_get_permission_result(NMClient *client, NMClientPermission permission)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
NMClientPermissionResult result = NM_CLIENT_PERMISSION_RESULT_UNKNOWN;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NM_CLIENT_PERMISSION_RESULT_UNKNOWN);
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
if (permission > NM_CLIENT_PERMISSION_NONE && permission <= NM_CLIENT_PERMISSION_LAST) {
priv = NM_CLIENT_GET_PRIVATE(client);
if (priv->permissions)
result = priv->permissions[permission - 1];
}
return result;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
/**
* nm_client_get_permissions_state:
* @self: the #NMClient instance
*
* Returns: the state of the cached permissions. %NM_TERNARY_DEFAULT
* means that no permissions result was yet received. All permissions
* are unknown. %NM_TERNARY_TRUE means that the permissions got received
* and are cached. %%NM_TERNARY_FALSE means that permissions are cached,
all: fix minor typos https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/565
2020-07-04 11:37:01 +03:00
* but they are invalided as "CheckPermissions" signal was received
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
* in the meantime.
*
* Since: 1.24
*/
NMTernary
nm_client_get_permissions_state(NMClient *self)
{
g_return_val_if_fail(NM_IS_CLIENT(self), NM_TERNARY_DEFAULT);
return NM_CLIENT_GET_PRIVATE(self)->permissions_state;
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_connectivity:
* @client: an #NMClient
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets the current network connectivity state. Contrast
* nm_client_check_connectivity() and
* nm_client_check_connectivity_async(), which re-check the
* connectivity state first before returning any information.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: the current connectivity state
*/
NMConnectivityState
nm_client_get_connectivity(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm,core: don't mix up enum types Touches a weak spot on clang's soul.
2015-03-18 18:08:24 +01:00
g_return_val_if_fail(NM_IS_CLIENT(client), NM_CONNECTIVITY_UNKNOWN);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->nm.connectivity;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_check_connectivity:
* @client: an #NMClient
* @cancellable: a #GCancellable
* @error: return location for a #GError
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Updates the network connectivity state and returns the (new)
* current state. Contrast nm_client_get_connectivity(), which returns
* the most recent known state without re-checking.
*
* This is a blocking call; use nm_client_check_connectivity_async()
* if you do not want to block.
*
* Returns: the (new) current connectivity state
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: Use nm_client_check_connectivity_async() or GDBusConnection.
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
*/
NMConnectivityState
nm_client_check_connectivity(NMClient *client, GCancellable *cancellable, GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
libnm: implement nm_client_check_connectivity() by using GDBusConnection directly Note that nm_client_check_connectivity() has this odd behavior of updating the connectivity state right away. We keep doing that.
2019-10-04 16:18:51 +02:00
gs_unref_variant GVariant *ret = NULL;
guint32 connectivity;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NM_CONNECTIVITY_UNKNOWN);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = _nm_client_dbus_call_sync(client,
libnm: implement nm_client_check_connectivity() by using GDBusConnection directly Note that nm_client_check_connectivity() has this odd behavior of updating the connectivity state right away. We keep doing that.
2019-10-04 16:18:51 +02:00
cancellable,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"CheckConnectivity",
g_variant_new("()"),
G_VARIANT_TYPE("(u)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
if (!ret)
libnm: clean up NMClient semantics when NM is not running Add a single function to check if NM is running and set a GError if not, then use it as appropriate. Don't bother to check if NM is running in get_*() functions if nm_running_changed_cb() would have reset the field anyway (and fix that up to reset a few more fields).
2014-10-07 09:26:40 -04:00
return NM_CONNECTIVITY_UNKNOWN;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_check_connectivity() by using GDBusConnection directly Note that nm_client_check_connectivity() has this odd behavior of updating the connectivity state right away. We keep doing that.
2019-10-04 16:18:51 +02:00
g_variant_get(ret, "(u)", &connectivity);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_check_connectivity() by using GDBusConnection directly Note that nm_client_check_connectivity() has this odd behavior of updating the connectivity state right away. We keep doing that.
2019-10-04 16:18:51 +02:00
/* upon receiving the synchronous response, we hack the NMClient state
* and update the property outside the ordered D-Bus messages (like
* "PropertiesChanged" signals).
*
* This is really ugly, we shouldn't do this. */
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv = NM_CLIENT_GET_PRIVATE(client);
if (priv->nm.connectivity != connectivity) {
priv->nm.connectivity = connectivity;
_notify(client, PROP_CONNECTIVITY);
}
libnm: implement nm_client_check_connectivity() by using GDBusConnection directly Note that nm_client_check_connectivity() has this odd behavior of updating the connectivity state right away. We keep doing that.
2019-10-04 16:18:51 +02:00
return connectivity;
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_check_connectivity_async:
* @client: an #NMClient
* @cancellable: a #GCancellable
* @callback: callback to call with the result
* @user_data: data for @callback.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Asynchronously updates the network connectivity state and invokes
* @callback when complete. Contrast nm_client_get_connectivity(),
* which (immediately) returns the most recent known state without
* re-checking, and nm_client_check_connectivity(), which blocks.
*/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_check_connectivity_async(NMClient *client,
GCancellable *cancellable,
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: implement nm_client_check_connectivity_async() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
g_return_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_check_connectivity_async() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
nm_client_check_connectivity_async,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"CheckConnectivity",
g_variant_new("()"),
G_VARIANT_TYPE("(u)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_variant_strip_dbus_error_cb);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
}
/**
* nm_client_check_connectivity_finish:
* @client: an #NMClient
* @result: the #GAsyncResult
* @error: return location for a #GError
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Retrieves the result of an nm_client_check_connectivity_async()
* call.
*
* Returns: the (new) current connectivity state
*/
NMConnectivityState
nm_client_check_connectivity_finish(NMClient *client, GAsyncResult *result, GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: implement nm_client_check_connectivity_async() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
gs_unref_variant GVariant *ret = NULL;
guint32 connectivity;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: clean up NMClient semantics when NM is not running Add a single function to check if NM is running and set a GError if not, then use it as appropriate. Don't bother to check if NM is running in get_*() functions if nm_running_changed_cb() would have reset the field anyway (and fix that up to reset a few more fields).
2014-10-07 09:26:40 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NM_CONNECTIVITY_UNKNOWN);
libnm: fix wrong assertion in nm_client_check_connectivity_finish() https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/476 Fixes: b44d883d920d ('libnm: implement nm_client_check_connectivity_async() by using GDBusConnection directly')
2020-06-25 14:03:57 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_check_connectivity_async),
NM_CONNECTIVITY_UNKNOWN);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
libnm: implement nm_client_check_connectivity_async() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
ret = g_task_propagate_pointer(G_TASK(result), error);
if (!ret)
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
return NM_CONNECTIVITY_UNKNOWN;
libnm: implement nm_client_check_connectivity_async() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
g_variant_get(ret, "(u)", &connectivity);
return connectivity;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
* nm_client_save_hostname:
* @client: the %NMClient
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @hostname: (nullable): the new persistent hostname to set, or %NULL to
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* clear any existing persistent hostname
* @cancellable: a #GCancellable, or %NULL
* @error: return location for #GError
*
* Requests that the machine's persistent hostname be set to the specified value
* or cleared.
*
* Returns: %TRUE if the request was successful, %FALSE if it failed
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: Use nm_client_save_hostname_async() or GDBusConnection.
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_save_hostname(NMClient *client,
const char *hostname,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
GCancellable *cancellable,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GError **error)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_save_hostname() by using GDBusConnection directly
2019-10-07 09:39:34 +02:00
g_return_val_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable), FALSE);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return _nm_client_dbus_call_sync_void(client,
libnm: implement nm_client_save_hostname() by using GDBusConnection directly
2019-10-07 09:39:34 +02:00
cancellable,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"SaveHostname",
g_variant_new("(s)", hostname ?: ""),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_save_hostname_async:
* @client: the %NMClient
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @hostname: (nullable): the new persistent hostname to set, or %NULL to
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* clear any existing persistent hostname
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the operation completes
* @user_data: caller-specific data passed to @callback
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
*
* Requests that the machine's persistent hostname be set to the specified value
* or cleared.
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_save_hostname_async(NMClient *client,
const char *hostname,
GCancellable *cancellable,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: implement nm_client_save_hostname_async() by using GDBusConnection directly
2019-10-07 09:43:47 +02:00
g_return_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_save_hostname_async() by using GDBusConnection directly
2019-10-07 09:43:47 +02:00
nm_client_save_hostname_async,
cancellable,
callback,
user_data,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"SaveHostname",
g_variant_new("(s)", hostname ?: ""),
G_VARIANT_TYPE("()"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_void_strip_dbus_error_cb);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_save_hostname_finish:
* @client: the %NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: return location for #GError
*
* Gets the result of an nm_client_save_hostname_async() call.
*
* Returns: %TRUE if the request was successful, %FALSE if it failed
**/
gboolean
nm_client_save_hostname_finish(NMClient *client, GAsyncResult *result, GError **error)
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_save_hostname_async() by using GDBusConnection directly
2019-10-07 09:43:47 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_save_hostname_async), FALSE);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: implement nm_client_save_hostname_async() by using GDBusConnection directly
2019-10-07 09:43:47 +02:00
return g_task_propagate_boolean(G_TASK(result), error);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
all: modify line separator comments to be 80 chars wide sed 's#^/\*\{5\}\*\+/$#/*****************************************************************************/#' $(git grep -l '\*\{5\}' | grep '\.[hc]$') -i
2016-10-02 18:22:50 +02:00
/*****************************************************************************/
libnm: implement nm_client_save_hostname_async() by using GDBusConnection directly
2019-10-07 09:43:47 +02:00
/* Devices */
all: modify line separator comments to be 80 chars wide sed 's#^/\*\{5\}\*\+/$#/*****************************************************************************/#' $(git grep -l '\*\{5\}' | grep '\.[hc]$') -i
2016-10-02 18:22:50 +02:00
/*****************************************************************************/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_devices:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets all the known network devices. Use nm_device_get_type() or the
docs: misc small fixes Cleans up all of the warnings that aren't overly annoying to clean up.
2014-12-18 12:44:57 -05:00
* <literal>NM_IS_DEVICE_XXXX</literal> functions to determine what kind of
* device member of the returned array is, and then you may use device-specific
* methods such as nm_device_ethernet_get_hw_address().
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: (transfer none) (element-type NMDevice): a #GPtrArray
* containing all the #NMDevices. The returned array is owned by the
* #NMClient object and should not be modified.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
const GPtrArray *
nm_client_get_devices(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return nml_dbus_property_ao_get_objs_as_ptrarray(
&NM_CLIENT_GET_PRIVATE(client)->nm.property_ao[PROPERTY_AO_IDX_DEVICES]);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
/**
* nm_client_get_all_devices:
* @client: a #NMClient
*
* Gets both real devices and device placeholders (eg, software devices which
* do not currently exist, but could be created automatically by NetworkManager
* if one of their NMDevice::ActivatableConnections was activated). Use
* nm_device_is_real() to determine whether each device is a real device or
* a placeholder.
*
* Use nm_device_get_type() or the NM_IS_DEVICE_XXXX() functions to determine
* what kind of device each member of the returned array is, and then you may
* use device-specific methods such as nm_device_ethernet_get_hw_address().
*
* Returns: (transfer none) (element-type NMDevice): a #GPtrArray
* containing all the #NMDevices. The returned array is owned by the
* #NMClient object and should not be modified.
*
* Since: 1.2
**/
const GPtrArray *
nm_client_get_all_devices(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return nml_dbus_property_ao_get_objs_as_ptrarray(
&NM_CLIENT_GET_PRIVATE(client)->nm.property_ao[PROPERTY_AO_IDX_ALL_DEVICES]);
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_device_by_path:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @object_path: the object path to search for
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets a #NMDevice from a #NMClient.
*
* Returns: (transfer none): the #NMDevice for the given @object_path or %NULL if none is found.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMDevice *
nm_client_get_device_by_path(NMClient *client, const char *object_path)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
g_return_val_if_fail(object_path, NULL);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return _dbobjs_get_nmobj_unpack_visible(client, object_path, NM_TYPE_DEVICE);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_device_by_iface:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @iface: the interface name to search for
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets a #NMDevice from a #NMClient.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: (transfer none): the #NMDevice for the given @iface or %NULL if none is found.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMDevice *
nm_client_get_device_by_iface(NMClient *client, const char *iface)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const GPtrArray *devices;
guint i;
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
g_return_val_if_fail(iface, NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
devices = nm_client_get_devices(client);
for (i = 0; i < devices->len; i++) {
NMDevice *candidate = g_ptr_array_index(devices, i);
if (nm_streq0(nm_device_get_iface(candidate), iface))
return candidate;
}
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NULL;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
all: modify line separator comments to be 80 chars wide sed 's#^/\*\{5\}\*\+/$#/*****************************************************************************/#' $(git grep -l '\*\{5\}' | grep '\.[hc]$') -i
2016-10-02 18:22:50 +02:00
/*****************************************************************************/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
/* Active Connections */
all: modify line separator comments to be 80 chars wide sed 's#^/\*\{5\}\*\+/$#/*****************************************************************************/#' $(git grep -l '\*\{5\}' | grep '\.[hc]$') -i
2016-10-02 18:22:50 +02:00
/*****************************************************************************/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_active_connections:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets the active connections.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: (transfer none) (element-type NMActiveConnection): a #GPtrArray
* containing all the active #NMActiveConnections.
* The returned array is owned by the client and should not be modified.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
const GPtrArray *
nm_client_get_active_connections(NMClient *client)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return nml_dbus_property_ao_get_objs_as_ptrarray(
&NM_CLIENT_GET_PRIVATE(client)->nm.property_ao[PROPERTY_AO_IDX_ACTIVE_CONNECTIONS]);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_get_primary_connection:
* @client: an #NMClient
*
* Gets the #NMActiveConnection corresponding to the primary active
* network device.
*
* In particular, when there is no VPN active, or the VPN does not
* have the default route, this returns the active connection that has
* the default route. If there is a VPN active with the default route,
* then this function returns the active connection that contains the
* route to the VPN endpoint.
*
* If there is no default route, or the default route is over a
* non-NetworkManager-recognized device, this will return %NULL.
*
* Returns: (transfer none): the appropriate #NMActiveConnection, if
* any
*/
NMActiveConnection *
nm_client_get_primary_connection(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return nml_dbus_property_o_get_obj(
&NM_CLIENT_GET_PRIVATE(client)->nm.property_o[PROPERTY_O_IDX_NM_PRIMAY_CONNECTION]);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
}
/**
* nm_client_get_activating_connection:
* @client: an #NMClient
*
* Gets the #NMActiveConnection corresponding to a
* currently-activating connection that is expected to become the new
* #NMClient:primary-connection upon successful activation.
*
* Returns: (transfer none): the appropriate #NMActiveConnection, if
* any.
*/
NMActiveConnection *
nm_client_get_activating_connection(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return nml_dbus_property_o_get_obj(
&NM_CLIENT_GET_PRIVATE(client)->nm.property_o[PROPERTY_O_IDX_NM_ACTIVATING_CONNECTION]);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
static void
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
activate_connection_cb(GObject *object, GAsyncResult *result, gpointer user_data)
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_unref_object GTask *task = user_data;
gs_unref_variant GVariant *ret = NULL;
const char *v_active_connection;
GError *error = NULL;
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(object), result, &error);
if (!ret) {
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_dbus_error_strip_remote_error(error);
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
g_task_return_error(task, error);
return;
}
libnm: split "NMManager" out of NMClient Clone NMClient as NMManager, and make NMClient just be a wrapper around that new class.
2014-09-24 13:14:48 -04:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
g_variant_get(ret, "(&o)", &v_active_connection);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_request_wait_start(g_steal_pointer(&task),
"ActivateConnection",
NM_TYPE_ACTIVE_CONNECTION,
v_active_connection,
NULL);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
}
/**
* nm_client_activate_connection_async:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @connection: (nullable): an #NMConnection
* @device: (nullable): the #NMDevice
* @specific_object: (nullable): the object path of a connection-type-specific
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* object this activation should use. This parameter is currently ignored for
* wired and mobile broadband connections, and the value of %NULL should be used
* (ie, no specific object). For Wi-Fi or WiMAX connections, pass the object
* path of a #NMAccessPoint or #NMWimaxNsp owned by @device, which you can
* get using nm_object_get_path(), and which will be used to complete the
* details of the newly added connection.
* @cancellable: a #GCancellable, or %NULL
* @callback: callback to be called when the activation has started
* @user_data: caller-specific data passed to @callback
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Asynchronously starts a connection to a particular network using the
* configuration settings from @connection and the network device @device.
* Certain connection types also take a "specific object" which is the object
* path of a connection- specific object, like an #NMAccessPoint for Wi-Fi
* connections, or an #NMWimaxNsp for WiMAX connections, to which you wish to
* connect. If the specific object is not given, NetworkManager can, in some
* cases, automatically determine which network to connect to given the settings
* in @connection.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* If @connection is not given for a device-based activation, NetworkManager
* picks the best available connection for the device and activates it.
*
* Note that the callback is invoked when NetworkManager has started activating
all/trivial: fix few comment typos
2017-05-11 12:08:02 +02:00
* the new connection, not when it finishes. You can use the returned
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* #NMActiveConnection object (in particular, #NMActiveConnection:state) to
* track the activation to its completion.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_activate_connection_async(NMClient *client,
NMConnection *connection,
NMDevice *device,
const char *specific_object,
GCancellable *cancellable,
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
const char *arg_connection = NULL;
const char *arg_device = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_if_fail(NM_IS_CLIENT(client));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
if (connection) {
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_if_fail(NM_IS_CONNECTION(connection));
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
arg_connection = nm_connection_get_path(connection);
g_return_if_fail(arg_connection);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
if (device) {
g_return_if_fail(NM_IS_DEVICE(device));
arg_device = nm_object_get_path(NM_OBJECT(device));
g_return_if_fail(arg_device);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_T(
client,
"ActivateConnection() for connection \"%s\", device \"%s\", specific_object \"%s",
arg_connection ?: "/",
arg_device ?: "/",
specific_object ?: "/");
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(
client,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
client,
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
nm_client_activate_connection_async,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"ActivateConnection",
g_variant_new("(ooo)", arg_connection ?: "/", arg_device ?: "/", specific_object ?: "/"),
G_VARIANT_TYPE("(o)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
activate_connection_cb);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_activate_connection_finish:
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets the result of a call to nm_client_activate_connection_async().
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: (transfer full): the new #NMActiveConnection on success, %NULL on
* failure, in which case @error will be set.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMActiveConnection *
nm_client_activate_connection_finish(NMClient *client, GAsyncResult *result, GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_ACTIVE_CONNECTION(
_request_wait_finish(client, result, nm_client_activate_connection_async, NULL, error));
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_add_and_activate_connection_done(GObject *object,
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
GAsyncResult *result,
gboolean use_add_and_activate_v2,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GTask *task_take)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_nm_unused gs_unref_object GTask *task = task_take;
gs_unref_variant GVariant *ret = NULL;
GError *error = NULL;
gs_unref_variant GVariant *v_result = NULL;
const char *v_active_connection;
const char *v_path;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(object), result, &error);
if (!ret) {
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_dbus_error_strip_remote_error(error);
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
g_task_return_error(task, error);
return;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
if (use_add_and_activate_v2) {
g_variant_get(ret, "(&o&o@a{sv})", &v_path, &v_active_connection, &v_result);
} else {
g_variant_get(ret, "(&o&o)", &v_path, &v_active_connection);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_request_wait_start(g_steal_pointer(&task),
"AddAndActivateConnection",
NM_TYPE_ACTIVE_CONNECTION,
v_active_connection,
g_steal_pointer(&v_result));
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
}
static void
_add_and_activate_connection_v1_cb(GObject *object, GAsyncResult *result, gpointer user_data)
{
_add_and_activate_connection_done(object, result, FALSE, user_data);
}
static void
_add_and_activate_connection_v2_cb(GObject *object, GAsyncResult *result, gpointer user_data)
{
_add_and_activate_connection_done(object, result, TRUE, user_data);
}
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_add_and_activate_connection(NMClient *self,
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
gboolean is_v2,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMConnection *partial,
NMDevice *device,
const char *specific_object,
GVariant *options,
GCancellable *cancellable,
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *arg_connection = NULL;
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
gboolean use_add_and_activate_v2 = FALSE;
const char *arg_device = NULL;
gpointer source_tag;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_if_fail(NM_IS_CLIENT(self));
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
g_return_if_fail(!partial || NM_IS_CONNECTION(partial));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
if (device) {
g_return_if_fail(NM_IS_DEVICE(device));
arg_device = nm_object_get_path(NM_OBJECT(device));
g_return_if_fail(arg_device);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
if (partial)
arg_connection = nm_connection_to_dbus(partial, NM_CONNECTION_SERIALIZE_ALL);
if (!arg_connection)
all: use nm_g_variant_singleton_a*() helpers
2021-04-14 21:12:02 +02:00
arg_connection = nm_g_variant_singleton_aLsaLsvII();
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
if (is_v2) {
if (!options)
all: use nm_g_variant_singleton_a*() helpers
2021-04-14 21:12:02 +02:00
options = nm_g_variant_singleton_aLsvI();
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
use_add_and_activate_v2 = TRUE;
source_tag = nm_client_add_and_activate_connection2;
} else {
if (options) {
if (g_variant_n_children(options) > 0)
use_add_and_activate_v2 = TRUE;
else
nm_clear_pointer(&options, nm_g_variant_unref_floating);
}
source_tag = nm_client_add_and_activate_connection_async;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "AddAndActivateConnection() started...");
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
if (use_add_and_activate_v2) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(self,
self,
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
source_tag,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"AddAndActivateConnection2",
g_variant_new("(@a{sa{sv}}oo@a{sv})",
arg_connection,
arg_device ?: "/",
specific_object ?: "/",
options),
G_VARIANT_TYPE("(ooa{sv})"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
_add_and_activate_connection_v2_cb);
} else {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(self,
self,
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
source_tag,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"AddAndActivateConnection",
g_variant_new("(@a{sa{sv}}oo)",
arg_connection,
arg_device ?: "/",
specific_object ?: "/"),
G_VARIANT_TYPE("(oo)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
_add_and_activate_connection_v1_cb);
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_add_and_activate_connection_async:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @partial: (nullable): an #NMConnection to add; the connection may be
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* partially filled (or even %NULL) and will be completed by NetworkManager
* using the given @device and @specific_object before being added
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @device: (nullable): the #NMDevice
* @specific_object: (nullable): the object path of a connection-type-specific
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* object this activation should use. This parameter is currently ignored for
* wired and mobile broadband connections, and the value of %NULL should be used
* (ie, no specific object). For Wi-Fi or WiMAX connections, pass the object
* path of a #NMAccessPoint or #NMWimaxNsp owned by @device, which you can
* get using nm_object_get_path(), and which will be used to complete the
* details of the newly added connection.
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
* If the variant is floating, it will be consumed.
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @cancellable: a #GCancellable, or %NULL
* @callback: callback to be called when the activation has started
* @user_data: caller-specific data passed to @callback
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Adds a new connection using the given details (if any) as a template,
* automatically filling in missing settings with the capabilities of the given
* device and specific object. The new connection is then asynchronously
* activated as with nm_client_activate_connection_async(). Cannot be used for
* VPN connections at this time.
libnm: return errors in nm_client_networking_set_enabled() and do not print an error in the library. The caller can decide what to do on an error.
2014-10-01 11:02:16 +02:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Note that the callback is invoked when NetworkManager has started activating
* the new connection, not when it finishes. You can used the returned
* #NMActiveConnection object (in particular, #NMActiveConnection:state) to
* track the activation to its completion.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_add_and_activate_connection_async(NMClient *client,
NMConnection *partial,
NMDevice *device,
const char *specific_object,
GCancellable *cancellable,
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
_add_and_activate_connection(client,
FALSE,
partial,
device,
specific_object,
NULL,
cancellable,
callback,
user_data);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_add_and_activate_connection_finish:
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets the result of a call to nm_client_add_and_activate_connection_async().
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* You can call nm_active_connection_get_connection() on the returned
* #NMActiveConnection to find the path of the created #NMConnection.
*
* Returns: (transfer full): the new #NMActiveConnection on success, %NULL on
* failure, in which case @error will be set.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMActiveConnection *
nm_client_add_and_activate_connection_finish(NMClient *client, GAsyncResult *result, GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_ACTIVE_CONNECTION(_request_wait_finish(client,
result,
nm_client_add_and_activate_connection_async,
NULL,
error));
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
}
/**
all: return output dictionary from "AddAndActivate2" Add a "a{sv}" output argument to "AddAndActivate2" D-Bus API. "AddAndActivate2" replaces "AddAndActivate" with more options. It also has a dictionary argument to be forward compatible so that we hopefully won't need an "AddAndActivate3". However, it lacked a similar output dictionary. Add it for future extensibility. I think this is really to workaround a shortcoming of D-Bus, which does provide strong typing and type information about its API, but does not allow to extend an existing API in a backward compatible manner. So we either resort to Method(), Method2(), Method3() variants, or a catch-all variant with a generic "a{sv}" input/output argument. In libnm, rename "nm_client_add_and_activate_connection_options()" to "nm_client_add_and_activate_connection2()". I think libnm API should have an obvious correspondence with D-Bus API. Or stated differently, if "AddAndActivateOptions" would be a better name, then the D-Bus API should be renamed. We should prefer one name over the other, but regardless of which is preferred, the naming for D-Bus and libnm API should correspond. In this case, I do think that AddAndActivate2() is a better name than AddAndActivateOptions(). Hence I rename the libnm API. Also, unless necessary, let libnm still call "AddAndActivate" instead of "AddAndActivate2". Our backward compatibility works the way that libnm requires a server version at least as new as itself. As such, libnm theoretically could assume that server version is new enough to support "AddAndActivate2" and could always use the more powerful variant. However, we don't need to break compatibility intentionally and for little gain. Here, it's easy to let libnm also handle old server API, by continuing to use "AddAndActivate" for nm_client_add_and_activate_connection(). Note that during package update, we don't restart the currently running NetworkManager instance. In such a scenario, it can easily happen that nmcli/libnm is newer than the server version. Let's try a bit harder to not break that. Changes as discussed in [1]. [1] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/37#note_79876
2018-12-20 07:48:31 +01:00
* nm_client_add_and_activate_connection2:
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
* @client: a #NMClient
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @partial: (nullable): an #NMConnection to add; the connection may be
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
* partially filled (or even %NULL) and will be completed by NetworkManager
* using the given @device and @specific_object before being added
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @device: (nullable): the #NMDevice
* @specific_object: (nullable): the object path of a connection-type-specific
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
* object this activation should use. This parameter is currently ignored for
* wired and mobile broadband connections, and the value of %NULL should be used
libnm/doc: fix spelling in nm_client_add_and_activate_connection2() documentation (cherry picked from commit c8a97031308f3308f7983eabce390e758f9cf11d)
2020-04-27 15:54:49 +02:00
* (i.e., no specific object). For Wi-Fi or WiMAX connections, pass the object
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
* path of a #NMAccessPoint or #NMWimaxNsp owned by @device, which you can
* get using nm_object_get_path(), and which will be used to complete the
* details of the newly added connection.
* @options: a #GVariant containing a dictionary with options, or %NULL
* @cancellable: a #GCancellable, or %NULL
* @callback: callback to be called when the activation has started
* @user_data: caller-specific data passed to @callback
*
* Adds a new connection using the given details (if any) as a template,
* automatically filling in missing settings with the capabilities of the given
* device and specific object. The new connection is then asynchronously
* activated as with nm_client_activate_connection_async(). Cannot be used for
* VPN connections at this time.
*
* Note that the callback is invoked when NetworkManager has started activating
* the new connection, not when it finishes. You can used the returned
* #NMActiveConnection object (in particular, #NMActiveConnection:state) to
* track the activation to its completion.
*
libnm/doc: fix spelling in nm_client_add_and_activate_connection2() documentation (cherry picked from commit c8a97031308f3308f7983eabce390e758f9cf11d)
2020-04-27 15:54:49 +02:00
* This is identical to nm_client_add_and_activate_connection_async() but takes
all: fix typo in man pages There should be a comma after 'Otherwise' and 'Currently'. https://bugzilla.redhat.com/show_bug.cgi?id=1852452 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/560
2020-07-01 17:20:40 -04:00
* a further @options parameter. Currently, the following options are supported
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
* by the daemon:
* * "persist": A string describing how the connection should be stored.
* The default is "disk", but it can be modified to "memory" (until
* the daemon quits) or "volatile" (will be deleted on disconnect).
all: rename "bind" option for AddAndActivateConnection2 to "bind-activation" "bind" specifically binds the lifetime of the activation (NMActiveConnection). In combination with "persist=volatile", the lifetime of the NMSettingsConnection is indirectly bound to the NMActiveConnection. But still these concepts make sense independently. In the future, it may make sense to also bind the lifetime of the NMSettingsConnection to the D-Bus client. Hence, rename the option to allow for the distinction. Also, belatedly fix libnm comment about "bind" only working with "persist" "volatile". Fixes: eb883e34a5889812d35e5c2d7bc6051723001a3a
2018-11-19 11:16:44 +01:00
* * "bind-activation": Bind the connection lifetime to something. The default is "none",
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
* meaning an explicit disconnect is needed. The value "dbus-client"
all: rename "bind" option for AddAndActivateConnection2 to "bind-activation" "bind" specifically binds the lifetime of the activation (NMActiveConnection). In combination with "persist=volatile", the lifetime of the NMSettingsConnection is indirectly bound to the NMActiveConnection. But still these concepts make sense independently. In the future, it may make sense to also bind the lifetime of the NMSettingsConnection to the D-Bus client. Hence, rename the option to allow for the distinction. Also, belatedly fix libnm comment about "bind" only working with "persist" "volatile". Fixes: eb883e34a5889812d35e5c2d7bc6051723001a3a
2018-11-19 11:16:44 +01:00
* means the connection will automatically be deactivated when the calling
libnm/doc: fix spelling in nm_client_add_and_activate_connection2() documentation (cherry picked from commit c8a97031308f3308f7983eabce390e758f9cf11d)
2020-04-27 15:54:49 +02:00
* D-Bus client disappears from the system bus.
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
*
* Since: 1.16
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_add_and_activate_connection2(NMClient *client,
NMConnection *partial,
NMDevice *device,
const char *specific_object,
GVariant *options,
GCancellable *cancellable,
all: return output dictionary from "AddAndActivate2" Add a "a{sv}" output argument to "AddAndActivate2" D-Bus API. "AddAndActivate2" replaces "AddAndActivate" with more options. It also has a dictionary argument to be forward compatible so that we hopefully won't need an "AddAndActivate3". However, it lacked a similar output dictionary. Add it for future extensibility. I think this is really to workaround a shortcoming of D-Bus, which does provide strong typing and type information about its API, but does not allow to extend an existing API in a backward compatible manner. So we either resort to Method(), Method2(), Method3() variants, or a catch-all variant with a generic "a{sv}" input/output argument. In libnm, rename "nm_client_add_and_activate_connection_options()" to "nm_client_add_and_activate_connection2()". I think libnm API should have an obvious correspondence with D-Bus API. Or stated differently, if "AddAndActivateOptions" would be a better name, then the D-Bus API should be renamed. We should prefer one name over the other, but regardless of which is preferred, the naming for D-Bus and libnm API should correspond. In this case, I do think that AddAndActivate2() is a better name than AddAndActivateOptions(). Hence I rename the libnm API. Also, unless necessary, let libnm still call "AddAndActivate" instead of "AddAndActivate2". Our backward compatibility works the way that libnm requires a server version at least as new as itself. As such, libnm theoretically could assume that server version is new enough to support "AddAndActivate2" and could always use the more powerful variant. However, we don't need to break compatibility intentionally and for little gain. Here, it's easy to let libnm also handle old server API, by continuing to use "AddAndActivate" for nm_client_add_and_activate_connection(). Note that during package update, we don't restart the currently running NetworkManager instance. In such a scenario, it can easily happen that nmcli/libnm is newer than the server version. Let's try a bit harder to not break that. Changes as discussed in [1]. [1] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/37#note_79876
2018-12-20 07:48:31 +01:00
GAsyncReadyCallback callback,
gpointer user_data)
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
{
libnm: implement nm_client_activate_connection()/nm_client_add_and_activate_connection*() by using GDBusConnection directly We still need the bits in "nm-manager.c", to wait until the NMActiveConnection instance is ready. This is now done by nm_manager_complete_active_connection().
2019-10-03 16:30:25 +02:00
_add_and_activate_connection(client,
TRUE,
partial,
device,
specific_object,
options,
cancellable,
callback,
user_data);
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
}
/**
all: return output dictionary from "AddAndActivate2" Add a "a{sv}" output argument to "AddAndActivate2" D-Bus API. "AddAndActivate2" replaces "AddAndActivate" with more options. It also has a dictionary argument to be forward compatible so that we hopefully won't need an "AddAndActivate3". However, it lacked a similar output dictionary. Add it for future extensibility. I think this is really to workaround a shortcoming of D-Bus, which does provide strong typing and type information about its API, but does not allow to extend an existing API in a backward compatible manner. So we either resort to Method(), Method2(), Method3() variants, or a catch-all variant with a generic "a{sv}" input/output argument. In libnm, rename "nm_client_add_and_activate_connection_options()" to "nm_client_add_and_activate_connection2()". I think libnm API should have an obvious correspondence with D-Bus API. Or stated differently, if "AddAndActivateOptions" would be a better name, then the D-Bus API should be renamed. We should prefer one name over the other, but regardless of which is preferred, the naming for D-Bus and libnm API should correspond. In this case, I do think that AddAndActivate2() is a better name than AddAndActivateOptions(). Hence I rename the libnm API. Also, unless necessary, let libnm still call "AddAndActivate" instead of "AddAndActivate2". Our backward compatibility works the way that libnm requires a server version at least as new as itself. As such, libnm theoretically could assume that server version is new enough to support "AddAndActivate2" and could always use the more powerful variant. However, we don't need to break compatibility intentionally and for little gain. Here, it's easy to let libnm also handle old server API, by continuing to use "AddAndActivate" for nm_client_add_and_activate_connection(). Note that during package update, we don't restart the currently running NetworkManager instance. In such a scenario, it can easily happen that nmcli/libnm is newer than the server version. Let's try a bit harder to not break that. Changes as discussed in [1]. [1] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/37#note_79876
2018-12-20 07:48:31 +01:00
* nm_client_add_and_activate_connection2_finish:
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
doc: reorder gtkdoc annotations for consistency
2023-03-02 12:18:30 +01:00
* @out_result: (out) (optional) (nullable) (transfer full): the output result
all: fix typo in man pages There should be a comma after 'Otherwise' and 'Currently'. https://bugzilla.redhat.com/show_bug.cgi?id=1852452 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/560
2020-07-01 17:20:40 -04:00
* of type "a{sv}" returned by D-Bus' AddAndActivate2 call. Currently, no
all: return output dictionary from "AddAndActivate2" Add a "a{sv}" output argument to "AddAndActivate2" D-Bus API. "AddAndActivate2" replaces "AddAndActivate" with more options. It also has a dictionary argument to be forward compatible so that we hopefully won't need an "AddAndActivate3". However, it lacked a similar output dictionary. Add it for future extensibility. I think this is really to workaround a shortcoming of D-Bus, which does provide strong typing and type information about its API, but does not allow to extend an existing API in a backward compatible manner. So we either resort to Method(), Method2(), Method3() variants, or a catch-all variant with a generic "a{sv}" input/output argument. In libnm, rename "nm_client_add_and_activate_connection_options()" to "nm_client_add_and_activate_connection2()". I think libnm API should have an obvious correspondence with D-Bus API. Or stated differently, if "AddAndActivateOptions" would be a better name, then the D-Bus API should be renamed. We should prefer one name over the other, but regardless of which is preferred, the naming for D-Bus and libnm API should correspond. In this case, I do think that AddAndActivate2() is a better name than AddAndActivateOptions(). Hence I rename the libnm API. Also, unless necessary, let libnm still call "AddAndActivate" instead of "AddAndActivate2". Our backward compatibility works the way that libnm requires a server version at least as new as itself. As such, libnm theoretically could assume that server version is new enough to support "AddAndActivate2" and could always use the more powerful variant. However, we don't need to break compatibility intentionally and for little gain. Here, it's easy to let libnm also handle old server API, by continuing to use "AddAndActivate" for nm_client_add_and_activate_connection(). Note that during package update, we don't restart the currently running NetworkManager instance. In such a scenario, it can easily happen that nmcli/libnm is newer than the server version. Let's try a bit harder to not break that. Changes as discussed in [1]. [1] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/37#note_79876
2018-12-20 07:48:31 +01:00
* output is implemented yet.
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
*
all: return output dictionary from "AddAndActivate2" Add a "a{sv}" output argument to "AddAndActivate2" D-Bus API. "AddAndActivate2" replaces "AddAndActivate" with more options. It also has a dictionary argument to be forward compatible so that we hopefully won't need an "AddAndActivate3". However, it lacked a similar output dictionary. Add it for future extensibility. I think this is really to workaround a shortcoming of D-Bus, which does provide strong typing and type information about its API, but does not allow to extend an existing API in a backward compatible manner. So we either resort to Method(), Method2(), Method3() variants, or a catch-all variant with a generic "a{sv}" input/output argument. In libnm, rename "nm_client_add_and_activate_connection_options()" to "nm_client_add_and_activate_connection2()". I think libnm API should have an obvious correspondence with D-Bus API. Or stated differently, if "AddAndActivateOptions" would be a better name, then the D-Bus API should be renamed. We should prefer one name over the other, but regardless of which is preferred, the naming for D-Bus and libnm API should correspond. In this case, I do think that AddAndActivate2() is a better name than AddAndActivateOptions(). Hence I rename the libnm API. Also, unless necessary, let libnm still call "AddAndActivate" instead of "AddAndActivate2". Our backward compatibility works the way that libnm requires a server version at least as new as itself. As such, libnm theoretically could assume that server version is new enough to support "AddAndActivate2" and could always use the more powerful variant. However, we don't need to break compatibility intentionally and for little gain. Here, it's easy to let libnm also handle old server API, by continuing to use "AddAndActivate" for nm_client_add_and_activate_connection(). Note that during package update, we don't restart the currently running NetworkManager instance. In such a scenario, it can easily happen that nmcli/libnm is newer than the server version. Let's try a bit harder to not break that. Changes as discussed in [1]. [1] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/37#note_79876
2018-12-20 07:48:31 +01:00
* Gets the result of a call to nm_client_add_and_activate_connection2().
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
*
* You can call nm_active_connection_get_connection() on the returned
* #NMActiveConnection to find the path of the created #NMConnection.
*
* Returns: (transfer full): the new #NMActiveConnection on success, %NULL on
* failure, in which case @error will be set.
libnm: fix a large amount of Since tags Some comments are malformed, some are missing altogether.
2022-11-07 21:13:28 +01:00
*
* Since: 1.16
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
**/
NMActiveConnection *
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_add_and_activate_connection2_finish(NMClient *client,
all: return output dictionary from "AddAndActivate2" Add a "a{sv}" output argument to "AddAndActivate2" D-Bus API. "AddAndActivate2" replaces "AddAndActivate" with more options. It also has a dictionary argument to be forward compatible so that we hopefully won't need an "AddAndActivate3". However, it lacked a similar output dictionary. Add it for future extensibility. I think this is really to workaround a shortcoming of D-Bus, which does provide strong typing and type information about its API, but does not allow to extend an existing API in a backward compatible manner. So we either resort to Method(), Method2(), Method3() variants, or a catch-all variant with a generic "a{sv}" input/output argument. In libnm, rename "nm_client_add_and_activate_connection_options()" to "nm_client_add_and_activate_connection2()". I think libnm API should have an obvious correspondence with D-Bus API. Or stated differently, if "AddAndActivateOptions" would be a better name, then the D-Bus API should be renamed. We should prefer one name over the other, but regardless of which is preferred, the naming for D-Bus and libnm API should correspond. In this case, I do think that AddAndActivate2() is a better name than AddAndActivateOptions(). Hence I rename the libnm API. Also, unless necessary, let libnm still call "AddAndActivate" instead of "AddAndActivate2". Our backward compatibility works the way that libnm requires a server version at least as new as itself. As such, libnm theoretically could assume that server version is new enough to support "AddAndActivate2" and could always use the more powerful variant. However, we don't need to break compatibility intentionally and for little gain. Here, it's easy to let libnm also handle old server API, by continuing to use "AddAndActivate" for nm_client_add_and_activate_connection(). Note that during package update, we don't restart the currently running NetworkManager instance. In such a scenario, it can easily happen that nmcli/libnm is newer than the server version. Let's try a bit harder to not break that. Changes as discussed in [1]. [1] https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/37#note_79876
2018-12-20 07:48:31 +01:00
GAsyncResult *result,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant **out_result,
GError **error)
libnm: Add support to pass options to AddAndActivateConnection This adds the new methods nm_client_add_and_activate_connection_options_* and ports the existing methods to use the new AddAndActivateConnection2 call rather than AddAndActivateConnection, allowing further parameters to be passed in.
2018-10-26 13:44:38 +02:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_ACTIVE_CONNECTION(_request_wait_finish(client,
result,
libnm: fix wrong assertion in nm_client_add_and_activate_connection2_finish() Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-01-15 12:19:35 +01:00
nm_client_add_and_activate_connection2,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
out_result,
error));
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_deactivate_connection:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @active: the #NMActiveConnection to deactivate
* @cancellable: a #GCancellable, or %NULL
* @error: location for a #GError, or %NULL
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Deactivates an active #NMActiveConnection.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: success or failure
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: Use nm_client_deactivate_connection_async() or GDBusConnection.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_deactivate_connection(NMClient *client,
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMActiveConnection *active,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: implement nm_client_deactivate_connection() by using GDBusConnection directly Eventually, I will refactor libnm to no longer use gdbus-codegen and no GDBusProxy. In preparation of that, we must stop using that API. As first step, change nm_client_deactivate_connection(). Note how this was done previously: - nm_client_deactivate_connection() calls nm_manager_deactivate_connection() - nmdbus_manager_call_deactivate_connection_sync() calls g_dbus_proxy_call_sync() - g_dbus_proxy_call_sync() calls g_dbus_connection_call_sync() Currently this is still a bit ugly, because NMClient doesn't directly track the GDBusConnection nor the name owner. Instead, we need to peel it out of the object manager. One day, that will all be nicer, but first get rid of gdbus-codegen.
2019-10-01 16:51:10 +02:00
const char *active_path;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_val_if_fail(NM_IS_ACTIVE_CONNECTION(active), FALSE);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: implement nm_client_deactivate_connection() by using GDBusConnection directly Eventually, I will refactor libnm to no longer use gdbus-codegen and no GDBusProxy. In preparation of that, we must stop using that API. As first step, change nm_client_deactivate_connection(). Note how this was done previously: - nm_client_deactivate_connection() calls nm_manager_deactivate_connection() - nmdbus_manager_call_deactivate_connection_sync() calls g_dbus_proxy_call_sync() - g_dbus_proxy_call_sync() calls g_dbus_connection_call_sync() Currently this is still a bit ugly, because NMClient doesn't directly track the GDBusConnection nor the name owner. Instead, we need to peel it out of the object manager. One day, that will all be nicer, but first get rid of gdbus-codegen.
2019-10-01 16:51:10 +02:00
active_path = nm_object_get_path(NM_OBJECT(active));
g_return_val_if_fail(active_path, FALSE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return _nm_client_dbus_call_sync_void(client,
libnm: implement nm_client_deactivate_connection() by using GDBusConnection directly Eventually, I will refactor libnm to no longer use gdbus-codegen and no GDBusProxy. In preparation of that, we must stop using that API. As first step, change nm_client_deactivate_connection(). Note how this was done previously: - nm_client_deactivate_connection() calls nm_manager_deactivate_connection() - nmdbus_manager_call_deactivate_connection_sync() calls g_dbus_proxy_call_sync() - g_dbus_proxy_call_sync() calls g_dbus_connection_call_sync() Currently this is still a bit ugly, because NMClient doesn't directly track the GDBusConnection nor the name owner. Instead, we need to peel it out of the object manager. One day, that will all be nicer, but first get rid of gdbus-codegen.
2019-10-01 16:51:10 +02:00
cancellable,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"DeactivateConnection",
g_variant_new("(o)", active_path),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
}
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_deactivate_connection_async:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
* @client: a #NMClient
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* @active: the #NMActiveConnection to deactivate
* @cancellable: a #GCancellable, or %NULL
* @callback: callback to be called when the deactivation has completed
* @user_data: caller-specific data passed to @callback
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Asynchronously deactivates an active #NMActiveConnection.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_deactivate_connection_async(NMClient *client,
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
NMActiveConnection *active,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: implement nm_client_deactivate_connection_async() by using GDBusConnection directly
2019-10-02 13:38:36 +02:00
const char *active_path;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
g_return_if_fail(NM_IS_CLIENT(client));
g_return_if_fail(NM_IS_ACTIVE_CONNECTION(active));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_deactivate_connection_async() by using GDBusConnection directly
2019-10-02 13:38:36 +02:00
active_path = nm_object_get_path(NM_OBJECT(active));
g_return_if_fail(active_path);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_deactivate_connection_async() by using GDBusConnection directly
2019-10-02 13:38:36 +02:00
nm_client_deactivate_connection_async,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"DeactivateConnection",
g_variant_new("(o)", active_path),
G_VARIANT_TYPE("()"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_void_strip_dbus_error_cb);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
/**
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* nm_client_deactivate_connection_finish:
* @client: a #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Gets the result of a call to nm_client_deactivate_connection_async().
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm: (trivial) reorder NMClient function declarations/code Rearrange the NMClient function declarations and the functions themselves, and group them into "general", "device", and "active connection" sections. No code changes, just moving things around.
2014-09-18 16:11:00 -04:00
* Returns: success or failure
**/
gboolean
nm_client_deactivate_connection_finish(NMClient *client, GAsyncResult *result, GError **error)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
libnm: clean up NMClient semantics when NM is not running Add a single function to check if NM is running and set a GError if not, then use it as appropriate. Don't bother to check if NM is running in get_*() functions if nm_running_changed_cb() would have reset the field anyway (and fix that up to reset a few more fields).
2014-10-07 09:26:40 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_deactivate_connection_async() by using GDBusConnection directly
2019-10-02 13:38:36 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_deactivate_connection_async),
FALSE);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: implement nm_client_deactivate_connection_async() by using GDBusConnection directly
2019-10-02 13:38:36 +02:00
return g_task_propagate_boolean(G_TASK(result), error);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
all: modify line separator comments to be 80 chars wide sed 's#^/\*\{5\}\*\+/$#/*****************************************************************************/#' $(git grep -l '\*\{5\}' | grep '\.[hc]$') -i
2016-10-02 18:22:50 +02:00
/*****************************************************************************/
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/* Connections */
all: modify line separator comments to be 80 chars wide sed 's#^/\*\{5\}\*\+/$#/*****************************************************************************/#' $(git grep -l '\*\{5\}' | grep '\.[hc]$') -i
2016-10-02 18:22:50 +02:00
/*****************************************************************************/
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
libnm: change GSList to GPtrArray in libnm methods libnm mostly used GPtrArrays in its APIs, except that arrays of connections were usually GSLists. Fix this and make them GPtrArrays too (and rename nm_client_list_connections() to nm_client_get_connections() to match everything else).
2014-10-22 12:32:46 -04:00
* nm_client_get_connections:
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* @client: the %NMClient
*
libnm: change GSList to GPtrArray in libnm methods libnm mostly used GPtrArrays in its APIs, except that arrays of connections were usually GSLists. Fix this and make them GPtrArrays too (and rename nm_client_list_connections() to nm_client_get_connections() to match everything else).
2014-10-22 12:32:46 -04:00
* Returns: (transfer none) (element-type NMRemoteConnection): an array
* containing all connections provided by the remote settings service. The
* returned array is owned by the #NMClient object and should not be modified.
libnm: add code comments to hint that NMConnection might not validate
2016-03-23 17:36:28 +01:00
*
* The connections are as received from D-Bus and might not validate according
* to nm_connection_verify().
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
libnm: change GSList to GPtrArray in libnm methods libnm mostly used GPtrArrays in its APIs, except that arrays of connections were usually GSLists. Fix this and make them GPtrArrays too (and rename nm_client_list_connections() to nm_client_get_connections() to match everything else).
2014-10-22 12:32:46 -04:00
const GPtrArray *
nm_client_get_connections(NMClient *client)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return nml_dbus_property_ao_get_objs_as_ptrarray(
&NM_CLIENT_GET_PRIVATE(client)->settings.connections);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_get_connection_by_id:
* @client: the %NMClient
* @id: the id of the remote connection
*
* Returns the first matching %NMRemoteConnection matching a given @id.
*
* Returns: (transfer none): the remote connection object on success, or %NULL if no
* matching object was found.
libnm: add code comments to hint that NMConnection might not validate
2016-03-23 17:36:28 +01:00
*
* The connection is as received from D-Bus and might not validate according
* to nm_connection_verify().
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
NMRemoteConnection *
nm_client_get_connection_by_id(NMClient *client, const char *id)
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const GPtrArray *arr;
guint i;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_val_if_fail(id, NULL);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
arr = nm_client_get_connections(client);
for (i = 0; i < arr->len; i++) {
NMRemoteConnection *c = NM_REMOTE_CONNECTION(arr->pdata[i]);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (nm_streq0(id, nm_connection_get_id(NM_CONNECTION(c))))
return c;
}
return NULL;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_get_connection_by_path:
* @client: the %NMClient
* @path: the D-Bus object path of the remote connection
*
* Returns the %NMRemoteConnection representing the connection at @path.
*
* Returns: (transfer none): the remote connection object on success, or %NULL if the object was
* not known
libnm: add code comments to hint that NMConnection might not validate
2016-03-23 17:36:28 +01:00
*
* The connection is as received from D-Bus and might not validate according
* to nm_connection_verify().
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
NMRemoteConnection *
nm_client_get_connection_by_path(NMClient *client, const char *path)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
g_return_val_if_fail(path != NULL, NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return _dbobjs_get_nmobj_unpack_visible(client, path, NM_TYPE_REMOTE_CONNECTION);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_get_connection_by_uuid:
* @client: the %NMClient
* @uuid: the UUID of the remote connection
*
* Returns the %NMRemoteConnection identified by @uuid.
*
* Returns: (transfer none): the remote connection object on success, or %NULL if the object was
* not known
libnm: add code comments to hint that NMConnection might not validate
2016-03-23 17:36:28 +01:00
*
* The connection is as received from D-Bus and might not validate according
* to nm_connection_verify().
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
NMRemoteConnection *
nm_client_get_connection_by_uuid(NMClient *client, const char *uuid)
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const GPtrArray *arr;
guint i;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_val_if_fail(uuid, NULL);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
arr = nm_client_get_connections(client);
for (i = 0; i < arr->len; i++) {
NMRemoteConnection *c = NM_REMOTE_CONNECTION(arr->pdata[i]);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (nm_streq0(uuid, nm_connection_get_uuid(NM_CONNECTION(c))))
return c;
}
return NULL;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_add_connection_cb(GObject *source,
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
GAsyncResult *result,
gboolean with_extra_arg,
gpointer user_data)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
{
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
gs_unref_variant GVariant *ret = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_unref_object GTask *task = user_data;
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
gs_unref_variant GVariant *v_result = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *v_path;
GError *error = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(source), result, &error);
if (!ret) {
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_dbus_error_strip_remote_error(error);
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
g_task_return_error(task, error);
return;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
if (with_extra_arg) {
g_variant_get(ret, "(&o@a{sv})", &v_path, &v_result);
} else {
g_variant_get(ret, "(&o)", &v_path);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_request_wait_start(g_steal_pointer(&task),
"AddConnection",
NM_TYPE_REMOTE_CONNECTION,
v_path,
g_steal_pointer(&v_result));
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
}
static void
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
_add_connection_cb_without_extra_result(GObject *object, GAsyncResult *result, gpointer user_data)
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
{
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
_add_connection_cb(object, result, FALSE, user_data);
}
static void
_add_connection_cb_with_extra_result(GObject *object, GAsyncResult *result, gpointer user_data)
{
_add_connection_cb(object, result, TRUE, user_data);
}
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_add_connection_call(NMClient *self,
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
gpointer source_tag,
gboolean ignore_out_result,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *settings,
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
NMSettingsAddConnection2Flags flags,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *args,
GCancellable *cancellable,
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(self));
g_return_if_fail(!settings || g_variant_is_of_type(settings, G_VARIANT_TYPE("a{sa{sv}}")));
g_return_if_fail(!args || g_variant_is_of_type(args, G_VARIANT_TYPE("a{sv}")));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "AddConnection() started...");
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
if (!settings)
all: use nm_g_variant_singleton_a*() helpers
2021-04-14 21:12:02 +02:00
settings = nm_g_variant_singleton_aLsaLsvII();
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
/* Although AddConnection2() being capable to handle also AddConnection() and
* AddConnectionUnsaved() variants, we prefer to use the old D-Bus methods when
* they are sufficient. The reason is that libnm should avoid hard dependencies
* on 1.20 API whenever possible. */
if (ignore_out_result && flags == NM_SETTINGS_ADD_CONNECTION2_FLAG_TO_DISK) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(self,
self,
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
source_tag,
cancellable,
callback,
user_data,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"AddConnection",
g_variant_new("(@a{sa{sv}})", settings),
G_VARIANT_TYPE("(o)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
_add_connection_cb_without_extra_result);
} else if (ignore_out_result && flags == NM_SETTINGS_ADD_CONNECTION2_FLAG_IN_MEMORY) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(self,
self,
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
source_tag,
cancellable,
callback,
user_data,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"AddConnectionUnsaved",
g_variant_new("(@a{sa{sv}})", settings),
G_VARIANT_TYPE("(o)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
_add_connection_cb_without_extra_result);
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
} else {
all: use nm_g_variant_singleton_a*() helpers
2021-04-14 21:12:02 +02:00
_nm_client_dbus_call(self,
self,
source_tag,
cancellable,
callback,
user_data,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"AddConnection2",
g_variant_new("(@a{sa{sv}}u@a{sv})",
settings,
(guint32) flags,
args ?: nm_g_variant_singleton_aLsvI()),
G_VARIANT_TYPE("(oa{sv})"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
_add_connection_cb_with_extra_result);
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
}
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
}
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
* nm_client_add_connection_async:
* @client: the %NMClient
* @connection: the connection to add. Note that this object's settings will be
* added, not the object itself
* @save_to_disk: whether to immediately save the connection to disk
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the add operation completes
* @user_data: caller-specific data passed to @callback
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
*
* Requests that the remote settings service add the given settings to a new
* connection. If @save_to_disk is %TRUE, the connection is immediately written
* to disk; otherwise it is initially only stored in memory, but may be saved
* later by calling the connection's nm_remote_connection_commit_changes()
* method.
*
* @connection is untouched by this function and only serves as a template of
* the settings to add. The #NMRemoteConnection object that represents what
* NetworkManager actually added is returned to @callback when the addition
* operation is complete.
*
* Note that the #NMRemoteConnection returned in @callback may not contain
* identical settings to @connection as NetworkManager may perform automatic
* completion and/or normalization of connection properties.
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_add_connection_async(NMClient *client,
NMConnection *connection,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
gboolean save_to_disk,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CONNECTION(connection));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
_add_connection_call(client,
nm_client_add_connection_async,
TRUE,
nm_connection_to_dbus(connection, NM_CONNECTION_SERIALIZE_ALL),
save_to_disk ? NM_SETTINGS_ADD_CONNECTION2_FLAG_TO_DISK
: NM_SETTINGS_ADD_CONNECTION2_FLAG_IN_MEMORY,
NULL,
cancellable,
callback,
user_data);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_add_connection_finish:
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
* Gets the result of a call to nm_client_add_connection_async().
*
* Returns: (transfer full): the new #NMRemoteConnection on success, %NULL on
* failure, in which case @error will be set.
**/
NMRemoteConnection *
nm_client_add_connection_finish(NMClient *client, GAsyncResult *result, GError **error)
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_REMOTE_CONNECTION(
_request_wait_finish(client, result, nm_client_add_connection_async, NULL, error));
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
}
/**
* nm_client_add_connection2:
* @client: the %NMClient
* @settings: the "a{sa{sv}}" #GVariant with the content of the setting.
* @flags: the %NMSettingsAddConnection2Flags argument.
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @args: (nullable): the "a{sv}" #GVariant with extra argument or %NULL
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
* for no extra arguments.
* @ignore_out_result: this function wraps AddConnection2(), which has an
* additional result "a{sv}" output parameter. By setting this to %TRUE,
* you signal that you are not interested in that output parameter.
* This allows the function to fall back to AddConnection() and AddConnectionUnsaved(),
* which is interesting if you run against an older server version that does
* not yet provide AddConnection2(). By setting this to %FALSE, the function
* under the hood always calls AddConnection2().
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the add operation completes
* @user_data: caller-specific data passed to @callback
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
*
* Call AddConnection2() D-Bus API asynchronously.
*
* Since: 1.20
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_add_connection2(NMClient *client,
GVariant *settings,
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
NMSettingsAddConnection2Flags flags,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *args,
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
gboolean ignore_out_result,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
{
libnm: implement nm_client_add_connection*() by using GDBusConnection directly
2019-10-07 09:48:09 +02:00
_add_connection_call(client,
nm_client_add_connection2,
ignore_out_result,
settings,
flags,
args,
cancellable,
callback,
user_data);
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
}
/**
* nm_client_add_connection2_finish:
* @client: the #NMClient
* @result: the #GAsyncResult
doc: reorder gtkdoc annotations for consistency
2023-03-02 12:18:30 +01:00
* @out_result: (out) (optional) (nullable) (transfer full): the output
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* #GVariant from AddConnection2().
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
* If you care about the output result, then the "ignore_out_result"
* parameter of nm_client_add_connection2() must not be set to %TRUE.
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @error: the error argument.
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
*
* Returns: (transfer full): on success, a pointer to the added
* #NMRemoteConnection.
*
* Since: 1.20
*/
NMRemoteConnection *
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_add_connection2_finish(NMClient *client,
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
GAsyncResult *result,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant **out_result,
GError **error)
core,libnm: add AddConnection2() D-Bus API to block autoconnect from the start It should be possible to add a profile with autoconnect blocked form the start. Update2() has a %NM_SETTINGS_UPDATE2_FLAG_BLOCK_AUTOCONNECT flag to block autoconnect, and so we need something similar when adding a connection. As the existing AddConnection() and AddConnectionUnsaved() API is not extensible, add AddConnection2() that has flags and room for additional arguments. Then add and implement the new flag %NM_SETTINGS_ADD_CONNECTION2_FLAG_BLOCK_AUTOCONNECT for AddConnection2(). Note that libnm's nm_client_add_connection2() API can completely replace the existing nm_client_add_connection_async() call. In particular, it will automatically prefer to call the D-Bus methods AddConnection() and AddConnectionUnsaved(), in order to work with server versions older than 1.20. The purpose of this is that when upgrading the package, the running NetworkManager might still be older than the installed libnm. Anyway, so since nm_client_add_connection2_finish() also has a result output, the caller needs to decide whether he cares about that result. Hence it has an argument ignore_out_result, which allows to fallback to the old API. One might argue that a caller who doesn't care about the output results while still wanting to be backward compatible, should itself choose to call nm_client_add_connection_async() or nm_client_add_connection2(). But instead, it's more convenient if the new function can fully replace the old one, so that the caller does not need to switch which start/finish method to call. https://bugzilla.redhat.com/show_bug.cgi?id=1677068
2019-07-09 15:22:01 +02:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_REMOTE_CONNECTION(
_request_wait_finish(client, result, nm_client_add_connection2, out_result, error));
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
* nm_client_load_connections:
* @client: the %NMClient
libnm: fix introspection annotations for nm_client_load_connections() Add the "(array zero-terminated=1)" GLib introspection annotation when the parameter is a NULL-terminated string array. https://mail.gnome.org/archives/networkmanager-list/2016-August/msg00017.html Reported-by: Petr Horacek <phoracek@redhat.com>
2016-08-04 13:38:24 +02:00
* @filenames: (array zero-terminated=1): %NULL-terminated array of filenames to load
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* @failures: (out) (transfer full): on return, a %NULL-terminated array of
* filenames that failed to load
* @cancellable: a #GCancellable, or %NULL
* @error: return location for #GError
*
* Requests that the remote settings service load or reload the given files,
* adding or updating the connections described within.
*
* The changes to the indicated files will not yet be reflected in
* @client's connections array when the function returns.
*
* If all of the indicated files were successfully loaded, the
* function will return %TRUE, and @failures will be set to %NULL. If
* NetworkManager tried to load the files, but some (or all) failed,
* then @failures will be set to a %NULL-terminated array of the
* filenames that failed to load.
*
libnm: fix return value for nm_remote_settings_load_connections() to ignore server result nm_remote_settings_load_connections() and nm_remote_settings_load_connections_async() behave inconsistently. It's unexpected, that a FALSE return value may leave @error unset. Note that before commit 22e830f0469a ('settings/d-bus: fix boolean return value of "LoadConnections"'), the server boolean response would have been bogus anyway (at least for some versions). Unify the behavior, and ignore the boolean return value.
2019-10-06 23:19:32 +02:00
* Returns: %TRUE on success.
*
* Warning: before libnm 1.22, the boolean return value was inconsistent.
* That is made worse, because when running against certain server versions
* before 1.20, the server would return wrong values for success/failure.
* This means, if you use this function in libnm before 1.22, you are advised
* to ignore the boolean return value and only look at @failures and @error.
* With libnm >= 1.22, the boolean return value corresponds to whether @error was
* set. Note that even in the success case, you might have individual @failures.
* With 1.22, the return value is consistent with nm_client_load_connections_finish().
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: Use nm_client_load_connections_async() or GDBusConnection.
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_load_connections(NMClient *client,
char **filenames,
char ***failures,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
GCancellable *cancellable,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GError **error)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
{
libnm: implement nm_client_load_connections() by using GDBusConnection directly
2019-10-06 22:50:29 +02:00
gs_unref_variant GVariant *ret = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_load_connections() by using GDBusConnection directly
2019-10-06 22:50:29 +02:00
g_return_val_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable), FALSE);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = _nm_client_dbus_call_sync(client,
libnm: implement nm_client_load_connections() by using GDBusConnection directly
2019-10-06 22:50:29 +02:00
cancellable,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"LoadConnections",
g_variant_new("(^as)", filenames ?: NM_PTRARRAY_EMPTY(char *)),
G_VARIANT_TYPE("(bas)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
if (!ret) {
*failures = NULL;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
return FALSE;
libnm: implement nm_client_load_connections() by using GDBusConnection directly
2019-10-06 22:50:29 +02:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: fix leak and return "failures" from nm_client_load_connections() Due to this, `nmcli connection load` would also not print a warning about failure to load obviously bogus files: $ nmcli connection load /bogus Note that load is also used to unload files, so if the file name is a possibly valid name for a non-existing file, there is no failure. For example, we get no warning for $ nmcli connection load /etc/NetworkManager/system-connections/bogus Even if currently no such file is loaded, then the operation would still silently succeed, instead of succeeding the first time only. That is because load should be idempotent. [thaller@redhat.com: rewrote commit message] Fixes: 4af6219226e0 ('libnm: implement nm_client_load_connections_async() by using GDBusConnection directly') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/794 https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/979
2021-09-13 11:00:13 +08:00
g_variant_get(ret, "(b^as)", NULL, failures);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_load_connections() by using GDBusConnection directly
2019-10-06 22:50:29 +02:00
return TRUE;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_load_connections_async:
* @client: the %NMClient
libnm: fix introspection annotations for nm_client_load_connections() Add the "(array zero-terminated=1)" GLib introspection annotation when the parameter is a NULL-terminated string array. https://mail.gnome.org/archives/networkmanager-list/2016-August/msg00017.html Reported-by: Petr Horacek <phoracek@redhat.com>
2016-08-04 13:38:24 +02:00
* @filenames: (array zero-terminated=1): %NULL-terminated array of filenames to load
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the operation completes
* @user_data: caller-specific data passed to @callback
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
*
* Requests that the remote settings service asynchronously load or reload the
* given files, adding or updating the connections described within.
*
* See nm_client_load_connections() for more details.
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_load_connections_async(NMClient *client,
char **filenames,
GCancellable *cancellable,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: implement nm_client_load_connections_async() by using GDBusConnection directly
2019-10-06 23:16:48 +02:00
g_return_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_load_connections_async() by using GDBusConnection directly
2019-10-06 23:16:48 +02:00
nm_client_load_connections_async,
cancellable,
callback,
user_data,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"LoadConnections",
g_variant_new("(^as)", filenames ?: NM_PTRARRAY_EMPTY(char *)),
G_VARIANT_TYPE("(bas)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_variant_strip_dbus_error_cb);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_load_connections_finish:
* @client: the %NMClient
libnm: fix introspection annotations for nm_client_load_connections() Add the "(array zero-terminated=1)" GLib introspection annotation when the parameter is a NULL-terminated string array. https://mail.gnome.org/archives/networkmanager-list/2016-August/msg00017.html Reported-by: Petr Horacek <phoracek@redhat.com>
2016-08-04 13:38:24 +02:00
* @failures: (out) (transfer full) (array zero-terminated=1): on return, a
* %NULL-terminated array of filenames that failed to load
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
* Gets the result of an nm_client_load_connections_async() call.
* See nm_client_load_connections() for more details.
*
libnm: fix return value for nm_remote_settings_load_connections() to ignore server result nm_remote_settings_load_connections() and nm_remote_settings_load_connections_async() behave inconsistently. It's unexpected, that a FALSE return value may leave @error unset. Note that before commit 22e830f0469a ('settings/d-bus: fix boolean return value of "LoadConnections"'), the server boolean response would have been bogus anyway (at least for some versions). Unify the behavior, and ignore the boolean return value.
2019-10-06 23:19:32 +02:00
* Returns: %TRUE on success.
* Note that even in the success case, you might have individual @failures.
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_load_connections_finish(NMClient *client,
char ***failures,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
GAsyncResult *result,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GError **error)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
{
libnm: implement nm_client_load_connections_async() by using GDBusConnection directly
2019-10-06 23:16:48 +02:00
gs_unref_variant GVariant *ret = NULL;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_load_connections_async() by using GDBusConnection directly
2019-10-06 23:16:48 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_load_connections_async),
FALSE);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: implement nm_client_load_connections_async() by using GDBusConnection directly
2019-10-06 23:16:48 +02:00
ret = g_task_propagate_pointer(G_TASK(result), error);
if (!ret) {
*failures = NULL;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
return FALSE;
libnm: implement nm_client_load_connections_async() by using GDBusConnection directly
2019-10-06 23:16:48 +02:00
}
g_variant_get(ret, "(b^as)", NULL, &failures);
libnm: fix return value for nm_remote_settings_load_connections() to ignore server result nm_remote_settings_load_connections() and nm_remote_settings_load_connections_async() behave inconsistently. It's unexpected, that a FALSE return value may leave @error unset. Note that before commit 22e830f0469a ('settings/d-bus: fix boolean return value of "LoadConnections"'), the server boolean response would have been bogus anyway (at least for some versions). Unify the behavior, and ignore the boolean return value.
2019-10-06 23:19:32 +02:00
return TRUE;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_reload_connections:
* @client: the #NMClient
* @cancellable: a #GCancellable, or %NULL
* @error: return location for #GError
*
* Requests that the remote settings service reload all connection
* files from disk, adding, updating, and removing connections until
* the in-memory state matches the on-disk state.
*
doc: use "Returns:" annotation instead of deprecated aliases
2023-03-02 11:59:17 +01:00
* Returns: %TRUE on success, %FALSE on failure
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm/doc: fix gtk-doc for deprecated markers in libnm
2020-03-13 19:45:09 +01:00
* Deprecated: 1.22: Use nm_client_reload_connections_async() or GDBusConnection.
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
gboolean
nm_client_reload_connections(NMClient *client, GCancellable *cancellable, GError **error)
{
libnm: implement nm_client_reload_connections() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
gs_unref_variant GVariant *ret = NULL;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_reload_connections() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
g_return_val_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable), FALSE);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
ret = _nm_client_dbus_call_sync(client,
libnm: implement nm_client_reload_connections() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
cancellable,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"ReloadConnections",
g_variant_new("()"),
G_VARIANT_TYPE("(b)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
TRUE,
error);
if (!ret)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
return FALSE;
libnm: implement nm_client_reload_connections() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
return TRUE;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_reload_connections_async:
* @client: the #NMClient
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the reload operation completes
* @user_data: caller-specific data passed to @callback
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
*
* Requests that the remote settings service begin reloading all connection
* files from disk, adding, updating, and removing connections until the
* in-memory state matches the on-disk state.
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_reload_connections_async(NMClient *client,
GCancellable *cancellable,
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: implement nm_client_reload_connections_async() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
g_return_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_reload_connections_async() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
nm_client_reload_connections_async,
cancellable,
callback,
user_data,
NM_DBUS_PATH_SETTINGS,
NM_DBUS_INTERFACE_SETTINGS,
"ReloadConnections",
g_variant_new("()"),
G_VARIANT_TYPE("(b)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_variant_strip_dbus_error_cb);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
/**
* nm_client_reload_connections_finish:
* @client: the #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: return location for #GError
*
* Gets the result of an nm_client_reload_connections_async() call.
*
doc: use "Returns:" annotation instead of deprecated aliases
2023-03-02 11:59:17 +01:00
* Returns: %TRUE on success, %FALSE on failure
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
**/
gboolean
nm_client_reload_connections_finish(NMClient *client, GAsyncResult *result, GError **error)
{
libnm: implement nm_client_reload_connections_async() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
gs_unref_variant GVariant *ret = NULL;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_reload_connections_async() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_reload_connections_async),
FALSE);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: implement nm_client_reload_connections_async() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
ret = g_task_propagate_pointer(G_TASK(result), error);
if (!ret)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
return FALSE;
libnm: implement nm_client_reload_connections_async() by using GDBusConnection directly
2019-10-07 09:30:29 +02:00
return TRUE;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
all: modify line separator comments to be 80 chars wide sed 's#^/\*\{5\}\*\+/$#/*****************************************************************************/#' $(git grep -l '\*\{5\}' | grep '\.[hc]$') -i
2016-10-02 18:22:50 +02:00
/*****************************************************************************/
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
/**
* nm_client_get_dns_mode:
* @client: the #NMClient
*
* Gets the current DNS processing mode.
*
doc: use "Returns:" annotation instead of deprecated aliases
2023-03-02 11:59:17 +01:00
* Returns: the DNS processing mode, or %NULL in case the
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
* value is not available.
*
* Since: 1.6
**/
const char *
nm_client_get_dns_mode(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->dns_manager.mode;
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
}
/**
* nm_client_get_dns_rc_manager:
* @client: the #NMClient
*
* Gets the current DNS resolv.conf manager.
*
doc: use "Returns:" annotation instead of deprecated aliases
2023-03-02 11:59:17 +01:00
* Returns: the resolv.conf manager or %NULL in case the
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
* value is not available.
*
* Since: 1.6
**/
const char *
nm_client_get_dns_rc_manager(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->dns_manager.rc_manager;
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
}
/**
* nm_client_get_dns_configuration:
* @client: a #NMClient
*
* Gets the current DNS configuration
*
* Returns: (transfer none) (element-type NMDnsEntry): a #GPtrArray
* containing #NMDnsEntry elements or %NULL in case the value is not
* available. The returned array is owned by the #NMClient object
* and should not be modified.
*
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
* Since: 1.6
**/
const GPtrArray *
nm_client_get_dns_configuration(NMClient *client)
libnm: split "NMManager" out of NMClient Clone NMClient as NMManager, and make NMClient just be a wrapper around that new class.
2014-09-24 13:14:48 -04:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CLIENT_GET_PRIVATE(client)->dns_manager.configuration;
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static NMLDBusNotifyUpdatePropFlags
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_notify_update_prop_dns_manager_configuration(NMClient *self,
NMLDBusObject *dbobj,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface *meta_iface,
guint dbus_property_idx,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *value)
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_ptrarray GPtrArray *configuration_old = NULL;
gs_unref_ptrarray GPtrArray *configuration_new = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(G_OBJECT(self) == dbobj->nmobj);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (value) {
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *entry_var_tmp;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GVariantIter iter;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GPtrArray *array;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
configuration_new = g_ptr_array_new_with_free_func((GDestroyNotify) nm_dns_entry_unref);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_iter_init(&iter, value);
while (g_variant_iter_next(&iter, "@a{sv}", &entry_var_tmp)) {
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_unref_variant GVariant *entry_var = entry_var_tmp;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_auto_free_variant_iter GVariantIter *iterp_nameservers = NULL;
nm_auto_free_variant_iter GVariantIter *iterp_domains = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_free char **nameservers = NULL;
gs_free char **domains = NULL;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gboolean vpn = FALSE;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMDnsEntry *entry;
char *interface = NULL;
char *str;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gint32 priority = 0;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!g_variant_lookup(entry_var, "nameservers", "as", &iterp_nameservers)
|| !g_variant_lookup(entry_var, "priority", "i", &priority)) {
g_warning("Ignoring invalid DNS configuration");
continue;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
array = g_ptr_array_new();
while (g_variant_iter_next(iterp_nameservers, "&s", &str))
g_ptr_array_add(array, str);
g_ptr_array_add(array, NULL);
nameservers = (char **) g_ptr_array_free(array, FALSE);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (g_variant_lookup(entry_var, "domains", "as", &iterp_domains)) {
array = g_ptr_array_new();
while (g_variant_iter_next(iterp_domains, "&s", &str))
g_ptr_array_add(array, str);
g_ptr_array_add(array, NULL);
domains = (char **) g_ptr_array_free(array, FALSE);
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_lookup(entry_var, "interface", "&s", &interface);
g_variant_lookup(entry_var, "vpn", "b", &vpn);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
entry = nm_dns_entry_new(interface,
(const char *const *) nameservers,
(const char *const *) domains,
priority,
vpn);
if (!entry) {
g_warning("Ignoring invalid DNS entry");
continue;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_ptr_array_add(configuration_new, entry);
}
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
configuration_old = priv->dns_manager.configuration;
priv->dns_manager.configuration = g_steal_pointer(&configuration_new);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NOTIFY;
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
}
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
/**
* nm_client_get_capabilities:
* @client: the #NMClient instance
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @length: (out) (optional): the number of returned capabilities.
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
*
* Returns: (transfer none) (array length=length): the
* list of capabilities reported by the server or %NULL
* if the capabilities are unknown.
* The numeric values correspond to #NMCapability enum.
* The array is terminated by a numeric zero sentinel
* at position @length.
*
* Since: 1.24
*/
const guint32 *
nm_client_get_capabilities(NMClient *client, gsize *length)
{
NMClientPrivate *priv;
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
priv = NM_CLIENT_GET_PRIVATE(client);
NM_SET_OUT(length, priv->nm.capabilities_len);
return priv->nm.capabilities_arr;
}
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
/**
* nm_client_get_version_info:
* @client: the #NMClient instance
* @length: (out): the number of returned capabilities.
*
* If available, the first element in the array is NM_VERSION which
* encodes the daemon version as "(major << 16 | minor << 8 | micro)".
libnmc: fix bug checking VersionInfo's capabilities Remove the `+ 31u` that was making that it would search for bit 1 at array's element 1, instead of element 0. Fixed comparison >len that shoudl be >=len. Fix a few typos. Fixes: bc6098d44106 ('libnm: add internal nmc_client_has_{version_info_v,version_info_capability,capability}() helper')
2024-12-04 14:24:38 +01:00
* The following elements are a bitfield of %NMVersionInfoCapability
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
* that indicate that the daemon supports a certain capability.
*
* Returns: (transfer none) (array length=length): the
* list of capabilities reported by the server or %NULL
* if the capabilities are unknown.
*
* Since: 1.42
*/
const guint32 *
nm_client_get_version_info(NMClient *client, gsize *length)
{
NMClientPrivate *priv;
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
g_return_val_if_fail(length, NULL);
priv = NM_CLIENT_GET_PRIVATE(client);
*length = priv->nm.version_info_len;
return priv->nm.version_info_arr;
}
static NMLDBusNotifyUpdatePropFlags
_notify_update_prop_nm_au(guint32 **p_arr, gsize *p_len, GVariant *value)
{
nm_clear_g_free(p_arr);
*p_len = 0;
if (value) {
const guint32 *arr;
gsize len;
arr = g_variant_get_fixed_array(value, &len, sizeof(guint32));
*p_len = len;
*p_arr = g_new(guint32, len + 1);
if (len > 0)
memcpy(*p_arr, arr, len * sizeof(guint32));
(*p_arr)[len] = 0;
}
return NML_DBUS_NOTIFY_UPDATE_PROP_FLAGS_NOTIFY;
}
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
static NMLDBusNotifyUpdatePropFlags
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
_notify_update_prop_nm_capabilities(NMClient *self,
NMLDBusObject *dbobj,
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
const NMLDBusMetaIface *meta_iface,
guint dbus_property_idx,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GVariant *value)
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
nm_assert(G_OBJECT(self) == dbobj->nmobj);
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
return _notify_update_prop_nm_au(&priv->nm.capabilities_arr, &priv->nm.capabilities_len, value);
}
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
static NMLDBusNotifyUpdatePropFlags
_notify_update_prop_nm_version_info(NMClient *self,
NMLDBusObject *dbobj,
const NMLDBusMetaIface *meta_iface,
guint dbus_property_idx,
GVariant *value)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
nm_assert(G_OBJECT(self) == dbobj->nmobj);
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
return _notify_update_prop_nm_au(&priv->nm.version_info_arr, &priv->nm.version_info_len, value);
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
/**
* nm_client_get_checkpoints:
* @client: a #NMClient
*
* Gets all the active checkpoints.
*
* Returns: (transfer none) (element-type NMCheckpoint): a #GPtrArray
* containing all the #NMCheckpoint. The returned array is owned by the
* #NMClient object and should not be modified.
*
* Since: 1.12
**/
const GPtrArray *
nm_client_get_checkpoints(NMClient *client)
{
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return nml_dbus_property_ao_get_objs_as_ptrarray(
&NM_CLIENT_GET_PRIVATE(client)->nm.property_ao[PROPERTY_AO_IDX_CHECKPOINTS]);
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
libnm: implement nm_client_checkpoint_create() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
static void
checkpoint_create_cb(GObject *object, GAsyncResult *result, gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
gs_unref_object GTask *task = user_data;
gs_unref_variant GVariant *ret = NULL;
const char *v_checkpoint_path;
GError *error = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_checkpoint_create() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(object), result, &error);
if (!ret) {
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_dbus_error_strip_remote_error(error);
libnm: implement nm_client_checkpoint_create() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
g_task_return_error(task, error);
return;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_get(ret, "(&o)", &v_checkpoint_path);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_request_wait_start(g_steal_pointer(&task),
"CheckpointCreate",
NM_TYPE_CHECKPOINT,
v_checkpoint_path,
NULL);
libnm: implement nm_client_checkpoint_create() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
}
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
/**
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* nm_client_checkpoint_create:
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
* @client: the %NMClient
* @devices: (element-type NMDevice): a list of devices for which a
* checkpoint should be created.
* @rollback_timeout: the rollback timeout in seconds
* @flags: creation flags
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the add operation completes
* @user_data: caller-specific data passed to @callback
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
*
* Creates a checkpoint of the current networking configuration
* for given interfaces. An empty @devices argument means all
* devices. If @rollback_timeout is not zero, a rollback is
* automatically performed after the given timeout.
*
* Since: 1.12
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_checkpoint_create(NMClient *client,
const GPtrArray *devices,
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
guint32 rollback_timeout,
NMCheckpointCreateFlags flags,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
{
libnm: implement nm_client_checkpoint_create() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
gs_free const char **paths = NULL;
guint i;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
g_return_if_fail(NM_IS_CLIENT(client));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement nm_client_checkpoint_create() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
if (devices && devices->len > 0) {
paths = g_new(const char *, devices->len + 1);
for (i = 0; i < devices->len; i++)
paths[i] = nm_object_get_path(NM_OBJECT(devices->pdata[i]));
paths[i] = NULL;
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(
client,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
client,
libnm: implement nm_client_checkpoint_create() by using GDBusConnection directly
2019-10-04 17:42:32 +02:00
nm_client_checkpoint_create,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"CheckpointCreate",
g_variant_new("(^aouu)", paths ?: NM_PTRARRAY_EMPTY(const char *), rollback_timeout, flags),
G_VARIANT_TYPE("(o)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
checkpoint_create_cb);
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
/**
* nm_client_checkpoint_create_finish:
* @client: the #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* Gets the result of a call to nm_client_checkpoint_create().
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
*
* Returns: (transfer full): the new #NMCheckpoint on success, %NULL on
* failure, in which case @error will be set.
*
* Since: 1.12
**/
NMCheckpoint *
nm_client_checkpoint_create_finish(NMClient *client, GAsyncResult *result, GError **error)
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return NM_CHECKPOINT(
_request_wait_finish(client, result, nm_client_checkpoint_create, NULL, error));
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
/**
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* nm_client_checkpoint_destroy:
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
* @client: the %NMClient
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* @checkpoint_path: the D-Bus path for the checkpoint
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the add operation completes
* @user_data: caller-specific data passed to @callback
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
*
* Destroys an existing checkpoint without performing a rollback.
*
* Since: 1.12
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_checkpoint_destroy(NMClient *client,
const char *checkpoint_path,
GCancellable *cancellable,
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
g_return_if_fail(checkpoint_path && checkpoint_path[0] == '/');
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_checkpoint_destroy() by using GDBusConnection directly
2019-10-06 10:52:42 +02:00
nm_client_checkpoint_destroy,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"CheckpointDestroy",
g_variant_new("(o)", checkpoint_path),
G_VARIANT_TYPE("()"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_void_strip_dbus_error_cb);
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
/**
* nm_client_checkpoint_destroy_finish:
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* Gets the result of a call to nm_client_checkpoint_destroy().
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
*
* Returns: %TRUE on success or %FALSE on failure, in which case
* @error will be set.
*
* Since: 1.12
**/
gboolean
nm_client_checkpoint_destroy_finish(NMClient *client, GAsyncResult *result, GError **error)
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_checkpoint_destroy() by using GDBusConnection directly
2019-10-06 10:52:42 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_checkpoint_destroy), FALSE);
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
libnm: implement nm_client_checkpoint_destroy() by using GDBusConnection directly
2019-10-06 10:52:42 +02:00
return g_task_propagate_boolean(G_TASK(result), error);
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
/**
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* nm_client_checkpoint_rollback:
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
* @client: the %NMClient
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* @checkpoint_path: the D-Bus path to the checkpoint
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the add operation completes
* @user_data: caller-specific data passed to @callback
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
*
* Performs the rollback of a checkpoint before the timeout is reached.
*
* Since: 1.12
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_checkpoint_rollback(NMClient *client,
const char *checkpoint_path,
GCancellable *cancellable,
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
g_return_if_fail(checkpoint_path && checkpoint_path[0] == '/');
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_checkpoint_rollback() by using GDBusConnection directly
2019-10-06 11:06:52 +02:00
nm_client_checkpoint_rollback,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"CheckpointRollback",
g_variant_new("(o)", checkpoint_path),
G_VARIANT_TYPE("(a{su})"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_variant_strip_dbus_error_cb);
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
/**
* nm_client_checkpoint_rollback_finish:
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
* Gets the result of a call to nm_client_checkpoint_rollback().
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
*
* Returns: (transfer full) (element-type utf8 guint32): an hash table of
* devices and results. Devices are represented by their original
* D-Bus path; each result is a #NMRollbackResult.
*
* Since: 1.12
**/
GHashTable *
nm_client_checkpoint_rollback_finish(NMClient *client, GAsyncResult *result, GError **error)
{
libnm: implement nm_client_checkpoint_rollback() by using GDBusConnection directly
2019-10-06 11:06:52 +02:00
gs_unref_variant GVariant *ret = NULL;
gs_unref_variant GVariant *v_result = NULL;
GVariantIter iter;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GHashTable *hash;
const char *path;
libnm: implement nm_client_checkpoint_rollback() by using GDBusConnection directly
2019-10-06 11:06:52 +02:00
guint32 r;
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
g_return_val_if_fail(NM_IS_CLIENT(client), NULL);
libnm: implement nm_client_checkpoint_rollback() by using GDBusConnection directly
2019-10-06 11:06:52 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_checkpoint_rollback), NULL);
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
libnm: implement nm_client_checkpoint_rollback() by using GDBusConnection directly
2019-10-06 11:06:52 +02:00
ret = g_task_propagate_pointer(G_TASK(result), error);
if (!ret)
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
return NULL;
libnm: implement nm_client_checkpoint_rollback() by using GDBusConnection directly
2019-10-06 11:06:52 +02:00
g_variant_get(ret, "(@a{su})", &v_result);
hash = g_hash_table_new_full(nm_str_hash, g_str_equal, g_free, NULL);
g_variant_iter_init(&iter, v_result);
while (g_variant_iter_next(&iter, "{&su}", &path, &r))
g_hash_table_insert(hash, g_strdup(path), GUINT_TO_POINTER(r));
return hash;
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
}
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
/**
* nm_client_checkpoint_adjust_rollback_timeout:
* @client: the %NMClient
* @checkpoint_path: a D-Bus path to a checkpoint
* @add_timeout: the timeout in seconds counting from now.
* Set to zero, to disable the timeout.
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the add operation completes
* @user_data: caller-specific data passed to @callback
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
*
* Resets the timeout for the checkpoint with path @checkpoint_path
* to @timeout_add.
*
* Since: 1.12
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_checkpoint_adjust_rollback_timeout(NMClient *client,
const char *checkpoint_path,
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
guint32 add_timeout,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
libnm: rework checkpoint API The libnm API fir checkpoints was only introduced with 1.11. It is not yet stable, so there is still time to adjust it. Note that this changes API/ABI of the development branch. Changes: - we only add async variants of the checkpoint functions. I believe that synchronous D-Bus methods are fundamentally flawed, because they mess up the ordering of events. Rename the async functions by removing the "_async" suffix. This matches glib style, for which the async form is also not specially marked. - for function that refere to a particular checkpoint (rollback and destroy), accept the D-Bus path as string, instead of an NMCheckpoint instance. This form is more flexible, because it allows to use the function without having a NMCheckpoint instance at hand. On the other hand, if one has a NMCheckpoint instance, he can trivially obtain the path to make the call.
2018-04-04 13:33:53 +02:00
g_return_if_fail(checkpoint_path && checkpoint_path[0] == '/');
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_checkpoint_adjust_rollback_timeout() by using GDBusConnection directly
2019-10-06 11:16:10 +02:00
nm_client_checkpoint_adjust_rollback_timeout,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"CheckpointAdjustRollbackTimeout",
g_variant_new("(ou)", checkpoint_path, add_timeout),
G_VARIANT_TYPE("()"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_void_strip_dbus_error_cb);
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
}
/**
* nm_client_checkpoint_adjust_rollback_timeout_finish:
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
* Gets the result of a call to nm_client_checkpoint_adjust_rollback_timeout().
*
* Returns: %TRUE on success or %FALSE on failure.
*
* Since: 1.12
**/
gboolean
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_checkpoint_adjust_rollback_timeout_finish(NMClient *client,
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
GAsyncResult *result,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GError **error)
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_checkpoint_adjust_rollback_timeout() by using GDBusConnection directly
2019-10-06 11:16:10 +02:00
g_return_val_if_fail(
nm_g_task_is_valid(result, client, nm_client_checkpoint_adjust_rollback_timeout),
FALSE);
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
libnm: implement nm_client_checkpoint_adjust_rollback_timeout() by using GDBusConnection directly
2019-10-06 11:16:10 +02:00
return g_task_propagate_boolean(G_TASK(result), error);
checkpoint: allow resetting the rollback timeout via D-Bus This allows to adjust the timeout of an existing checkpoint. The main usecase of checkpoints, is to have a fail-safe when configuring the network remotely. By allowing to reset the timeout, the user can perform a series of actions, and keep bumping the timeout. That way, the entire series is still guarded by the same checkpoint, but the user can start with short timeout, and re-adjust the timeout as he goes along. The libnm API only implements the async form (at least for now). Sync methods are fundamentally wrong with D-Bus, and it's probably not needed. Also, follow glib convenction, where the async form doesn't have the _async name suffix. Also, accept a D-Bus path as argument, not a NMCheckpoint instance. The libnm API should not be more restricted than the underlying D-Bus API. It would be cumbersome to require the user to lookup the NMCheckpoint instance first, especially since libnm doesn't provide an efficient or convenient lookup-by-path method. On the other hand, retrieving the path from a NMCheckpoint instance is always possible.
2018-03-28 08:09:56 +02:00
}
libnm: add nm_client_reload() Introduce libnm API to reload NM configuration through the Reload() D-Bus method.
2019-09-05 15:19:33 +02:00
/**
* nm_client_reload:
* @client: the %NMClient
* @flags: flags indicating what to reload.
* @cancellable: a #GCancellable, or %NULL
libnm: fix warnings due to invalid "closure" annotation The "closure" annotation needs to be set on the callback parameter instead of on the data for the callback function. This patch fixes the following warning: """ ../src/libnm-core-impl/nm-utils.c:3632: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:4778: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5776: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5849: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:5976: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6091: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6448: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6521: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6581: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6663: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-client.c:6728: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:974: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1014: Warning: NM: invalid "closure" annotation: only valid on callback parameters ../src/libnm-client-impl/nm-secret-agent-old.c:1041: Warning: NM: invalid "closure" annotation: only valid on callback parameters """
2024-11-27 12:51:24 +01:00
* @callback: (scope async) (closure user_data): callback to be called when the add operation completes
* @user_data: caller-specific data passed to @callback
libnm: add nm_client_reload() Introduce libnm API to reload NM configuration through the Reload() D-Bus method.
2019-09-05 15:19:33 +02:00
*
* Reload NetworkManager's configuration and perform certain updates, like
* flushing caches or rewriting external state to disk. This is similar to
* sending SIGHUP to NetworkManager but it allows for more fine-grained control
* over what to reload (see @flags). It also allows non-root access via
* PolicyKit and contrary to signals it is synchronous.
*
* Since: 1.22
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_reload(NMClient *client,
libnm: add nm_client_reload() Introduce libnm API to reload NM configuration through the Reload() D-Bus method.
2019-09-05 15:19:33 +02:00
NMManagerReloadFlags flags,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: add nm_client_reload() Introduce libnm API to reload NM configuration through the Reload() D-Bus method.
2019-09-05 15:19:33 +02:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nm_client_dbus_call(client,
client,
libnm: implement nm_client_reload() by using GDBusConnection directly
2019-10-06 11:26:09 +02:00
nm_client_reload,
cancellable,
callback,
user_data,
NM_DBUS_PATH,
NM_DBUS_INTERFACE,
"Reload",
g_variant_new("(u)", (guint32) flags),
G_VARIANT_TYPE("()"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
nm_dbus_connection_call_finish_void_strip_dbus_error_cb);
libnm: add nm_client_reload() Introduce libnm API to reload NM configuration through the Reload() D-Bus method.
2019-09-05 15:19:33 +02:00
}
/**
* nm_client_reload_finish:
* @client: an #NMClient
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
* Gets the result of a call to nm_client_reload().
*
* Returns: %TRUE on success or %FALSE on failure.
*
* Since: 1.22
**/
gboolean
nm_client_reload_finish(NMClient *client, GAsyncResult *result, GError **error)
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
libnm: implement nm_client_reload() by using GDBusConnection directly
2019-10-06 11:26:09 +02:00
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_reload), FALSE);
libnm: add nm_client_reload() Introduce libnm API to reload NM configuration through the Reload() D-Bus method.
2019-09-05 15:19:33 +02:00
libnm: implement nm_client_reload() by using GDBusConnection directly
2019-10-06 11:26:09 +02:00
return g_task_propagate_boolean(G_TASK(result), error);
libnm: add nm_client_reload() Introduce libnm API to reload NM configuration through the Reload() D-Bus method.
2019-09-05 15:19:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: add nm_client_dbus_call() API Add an API for calling D-Bus methods arbitrary objects of NetworkManager's API. Of course, this is basically just a call to g_dbus_connection_call(), using the current name owner, nm_client_get_dbus_connection() and nm_client_get_main_context(). All of this could also be achieved without this new API. However, nm_client_dbus_call() also gracefully handles if the current name owner is %NULL. It's a valid concern whether such API is useful, as the users already have all pieces to do it themself. I think it is.
2020-03-13 17:30:23 +01:00
/**
* nm_client_dbus_call:
* @client: the #NMClient
* @object_path: path of remote object
* @interface_name: D-Bus interface to invoke method on
* @method_name: the name of the method to invoke
* @parameters: (nullable): a #GVariant tuple with parameters for the method
* or %NULL if not passing parameters
* @reply_type: (nullable): the expected type of the reply (which will be a
* tuple), or %NULL
* @timeout_msec: the timeout in milliseconds, -1 to use the default
* timeout or %G_MAXINT for no timeout
* @cancellable: (nullable): a #GCancellable or %NULL
* @callback: (nullable): a #GAsyncReadyCallback to call when the request
* is satisfied or %NULL if you don't care about the result of the
* method invocation
* @user_data: the data to pass to @callback
*
* Call g_dbus_connection_call() on the current name owner with the specified
* arguments. Most importantly, this invokes g_dbus_connection_call() with the
* client's #GMainContext, so that the response is always in order with other
* events D-Bus events. Of course, the call uses #GTask and will invoke the
* callback on the current g_main_context_get_thread_default().
*
* This API is merely a convenient wrapper for g_dbus_connection_call(). You can
* also use g_dbus_connection_call() directly, with the same effect.
*
* Since: 1.24
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_dbus_call(NMClient *client,
const char *object_path,
const char *interface_name,
const char *method_name,
GVariant *parameters,
libnm: add nm_client_dbus_call() API Add an API for calling D-Bus methods arbitrary objects of NetworkManager's API. Of course, this is basically just a call to g_dbus_connection_call(), using the current name owner, nm_client_get_dbus_connection() and nm_client_get_main_context(). All of this could also be achieved without this new API. However, nm_client_dbus_call() also gracefully handles if the current name owner is %NULL. It's a valid concern whether such API is useful, as the users already have all pieces to do it themself. I think it is.
2020-03-13 17:30:23 +01:00
const GVariantType *reply_type,
int timeout_msec,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: add nm_client_dbus_call() API Add an API for calling D-Bus methods arbitrary objects of NetworkManager's API. Of course, this is basically just a call to g_dbus_connection_call(), using the current name owner, nm_client_get_dbus_connection() and nm_client_get_main_context(). All of this could also be achieved without this new API. However, nm_client_dbus_call() also gracefully handles if the current name owner is %NULL. It's a valid concern whether such API is useful, as the users already have all pieces to do it themself. I think it is.
2020-03-13 17:30:23 +01:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add nm_client_dbus_call() API Add an API for calling D-Bus methods arbitrary objects of NetworkManager's API. Of course, this is basically just a call to g_dbus_connection_call(), using the current name owner, nm_client_get_dbus_connection() and nm_client_get_main_context(). All of this could also be achieved without this new API. However, nm_client_dbus_call() also gracefully handles if the current name owner is %NULL. It's a valid concern whether such API is useful, as the users already have all pieces to do it themself. I think it is.
2020-03-13 17:30:23 +01:00
_nm_client_dbus_call(client,
client,
nm_client_dbus_call,
cancellable,
callback,
user_data,
object_path,
interface_name,
method_name,
parameters,
reply_type,
G_DBUS_CALL_FLAGS_NONE,
timeout_msec == -1 ? NM_DBUS_DEFAULT_TIMEOUT_MSEC : timeout_msec,
nm_dbus_connection_call_finish_variant_cb);
}
/**
* nm_client_dbus_call_finish:
* @client: the #NMClient instance
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
* Gets the result of a call to nm_client_dbus_call().
*
* Returns: (transfer full): the result #GVariant or %NULL on error.
*
* Since: 1.24
**/
GVariant *
nm_client_dbus_call_finish(NMClient *client, GAsyncResult *result, GError **error)
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_dbus_call), FALSE);
return g_task_propagate_pointer(G_TASK(result), error);
}
/*****************************************************************************/
libnm: add nm_client_dbus_set_property() API Similar to nm_client_dbus_call(), but useful for setting a D-Bus property on NetworkManager's D-Bus interface. Note that we currently have various synchronous API for setting D-Bus properties (like nm_client_networking_set_enabled()). Synchronous API does not play well with the content of NMClient's cache, and was thus deprecated. However, until now no async variant exists. Instead of adding multiple async operations, I think it should be sufficient to only add one nm_client_dbus_set_property() property. It's still reasonably convenient to use for setting a property.
2020-03-13 18:11:56 +01:00
/**
* nm_client_dbus_set_property:
* @client: the #NMClient
* @object_path: path of remote object
libnm: fix documentation for value argument of nm_client_dbus_call() https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/451
2020-05-22 08:48:50 +02:00
* @interface_name: D-Bus interface for the property to set.
libnm: add nm_client_dbus_set_property() API Similar to nm_client_dbus_call(), but useful for setting a D-Bus property on NetworkManager's D-Bus interface. Note that we currently have various synchronous API for setting D-Bus properties (like nm_client_networking_set_enabled()). Synchronous API does not play well with the content of NMClient's cache, and was thus deprecated. However, until now no async variant exists. Instead of adding multiple async operations, I think it should be sufficient to only add one nm_client_dbus_set_property() property. It's still reasonably convenient to use for setting a property.
2020-03-13 18:11:56 +01:00
* @property_name: the name of the property to set
libnm: fix documentation for value argument of nm_client_dbus_call() https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/issues/451
2020-05-22 08:48:50 +02:00
* @value: a #GVariant with the value to set.
libnm: add nm_client_dbus_set_property() API Similar to nm_client_dbus_call(), but useful for setting a D-Bus property on NetworkManager's D-Bus interface. Note that we currently have various synchronous API for setting D-Bus properties (like nm_client_networking_set_enabled()). Synchronous API does not play well with the content of NMClient's cache, and was thus deprecated. However, until now no async variant exists. Instead of adding multiple async operations, I think it should be sufficient to only add one nm_client_dbus_set_property() property. It's still reasonably convenient to use for setting a property.
2020-03-13 18:11:56 +01:00
* @timeout_msec: the timeout in milliseconds, -1 to use the default
* timeout or %G_MAXINT for no timeout
* @cancellable: (nullable): a #GCancellable or %NULL
* @callback: (nullable): a #GAsyncReadyCallback to call when the request
* is satisfied or %NULL if you don't care about the result of the
* method invocation
* @user_data: the data to pass to @callback
*
* Like nm_client_dbus_call() but calls "Set" on the standard "org.freedesktop.DBus.Properties"
* D-Bus interface.
*
* Since: 1.24
**/
void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_client_dbus_set_property(NMClient *client,
const char *object_path,
const char *interface_name,
const char *property_name,
GVariant *value,
libnm: add nm_client_dbus_set_property() API Similar to nm_client_dbus_call(), but useful for setting a D-Bus property on NetworkManager's D-Bus interface. Note that we currently have various synchronous API for setting D-Bus properties (like nm_client_networking_set_enabled()). Synchronous API does not play well with the content of NMClient's cache, and was thus deprecated. However, until now no async variant exists. Instead of adding multiple async operations, I think it should be sufficient to only add one nm_client_dbus_set_property() property. It's still reasonably convenient to use for setting a property.
2020-03-13 18:11:56 +01:00
int timeout_msec,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: add nm_client_dbus_set_property() API Similar to nm_client_dbus_call(), but useful for setting a D-Bus property on NetworkManager's D-Bus interface. Note that we currently have various synchronous API for setting D-Bus properties (like nm_client_networking_set_enabled()). Synchronous API does not play well with the content of NMClient's cache, and was thus deprecated. However, until now no async variant exists. Instead of adding multiple async operations, I think it should be sufficient to only add one nm_client_dbus_set_property() property. It's still reasonably convenient to use for setting a property.
2020-03-13 18:11:56 +01:00
GAsyncReadyCallback callback,
gpointer user_data)
{
g_return_if_fail(NM_IS_CLIENT(client));
g_return_if_fail(interface_name);
g_return_if_fail(property_name);
g_return_if_fail(value);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add nm_client_dbus_set_property() API Similar to nm_client_dbus_call(), but useful for setting a D-Bus property on NetworkManager's D-Bus interface. Note that we currently have various synchronous API for setting D-Bus properties (like nm_client_networking_set_enabled()). Synchronous API does not play well with the content of NMClient's cache, and was thus deprecated. However, until now no async variant exists. Instead of adding multiple async operations, I think it should be sufficient to only add one nm_client_dbus_set_property() property. It's still reasonably convenient to use for setting a property.
2020-03-13 18:11:56 +01:00
_nm_client_dbus_call(client,
client,
nm_client_dbus_set_property,
cancellable,
callback,
user_data,
object_path,
DBUS_INTERFACE_PROPERTIES,
"Set",
g_variant_new("(ssv)", interface_name, property_name, value),
G_VARIANT_TYPE("()"),
G_DBUS_CALL_FLAGS_NONE,
timeout_msec == -1 ? NM_DBUS_DEFAULT_TIMEOUT_MSEC : timeout_msec,
nm_dbus_connection_call_finish_void_cb);
}
/**
* nm_client_dbus_set_property_finish:
* @client: the #NMClient instance
* @result: the result passed to the #GAsyncReadyCallback
* @error: location for a #GError, or %NULL
*
* Gets the result of a call to nm_client_dbus_set_property().
*
* Returns: %TRUE on success or %FALSE on failure.
*
* Since: 1.24
**/
gboolean
nm_client_dbus_set_property_finish(NMClient *client, GAsyncResult *result, GError **error)
{
g_return_val_if_fail(NM_IS_CLIENT(client), FALSE);
g_return_val_if_fail(nm_g_task_is_valid(result, client, nm_client_dbus_set_property), FALSE);
return g_task_propagate_boolean(G_TASK(result), error);
}
/*****************************************************************************/
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static void
_init_fetch_all(NMClient *self)
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_auto_pop_gmaincontext GMainContext *dbus_context = NULL;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_context = nm_g_main_context_push_thread_default_if_necessary(priv->dbus_context);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "fetch all");
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(!priv->get_managed_objects_cancellable);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->get_managed_objects_cancellable = g_cancellable_new();
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
shared: remove nm_dbus_connection_signal_subscribe_object_manager() helper It seems to complicate things more than helping. Drop it. What we still have is a wrapper around plain g_dbus_connection_signal_subscribe(). That one is trivial and helpful. The previous wrapper seems to add more complexity.
2020-01-14 12:38:07 +01:00
priv->dbsid_nm_object_manager =
nm_dbus_connection_signal_subscribe_object_manager(priv->dbus_connection,
priv->name_owner,
"/org/freedesktop",
NULL,
_dbus_managed_objects_changed_cb,
self,
NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->dbsid_dbus_properties_properties_changed =
nm_dbus_connection_signal_subscribe_properties_changed(priv->dbus_connection,
priv->name_owner,
NULL,
NULL,
_dbus_properties_changed_cb,
self,
NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->dbsid_nm_settings_connection_updated =
g_dbus_connection_signal_subscribe(priv->dbus_connection,
priv->name_owner,
NM_DBUS_INTERFACE_SETTINGS_CONNECTION,
"Updated",
NULL,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
NULL,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
G_DBUS_SIGNAL_FLAGS_NONE,
_dbus_settings_updated_cb,
self,
NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->dbsid_nm_connection_active_state_changed =
g_dbus_connection_signal_subscribe(priv->dbus_connection,
priv->name_owner,
NM_DBUS_INTERFACE_ACTIVE_CONNECTION,
"StateChanged",
NULL,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
NULL,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
G_DBUS_SIGNAL_FLAGS_NONE,
_dbus_nm_connection_active_state_changed_cb,
self,
NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->dbsid_nm_vpn_connection_state_changed =
g_dbus_connection_signal_subscribe(priv->dbus_connection,
priv->name_owner,
NM_DBUS_INTERFACE_VPN_CONNECTION,
"VpnStateChanged",
NULL,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
NULL,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
G_DBUS_SIGNAL_FLAGS_NONE,
_dbus_nm_vpn_connection_state_changed_cb,
self,
NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->dbsid_nm_check_permissions =
g_dbus_connection_signal_subscribe(priv->dbus_connection,
priv->name_owner,
NM_DBUS_INTERFACE,
"CheckPermissions",
NULL,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
NULL,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
G_DBUS_SIGNAL_FLAGS_NONE,
_dbus_nm_check_permissions_cb,
self,
NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
g_dbus_connection_call(priv->dbus_connection,
priv->name_owner,
"/org/freedesktop",
DBUS_INTERFACE_OBJECT_MANAGER,
"GetManagedObjects",
NULL,
G_VARIANT_TYPE("(a{oa{sa{sv}}})"),
G_DBUS_CALL_FLAGS_NO_AUTO_START,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
priv->get_managed_objects_cancellable,
_dbus_get_managed_objects_cb,
nm_utils_user_data_pack(self, g_object_ref(priv->context_busy_watcher)));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_dbus_check_permissions_start(self);
}
static void
_init_release_all(NMClient *self)
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
CList **dbus_objects_lst_heads;
NMLDBusObject *dbobj;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
int i;
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
gboolean permissions_state_changed = FALSE;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "release all");
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_cancellable(&priv->permissions_cancellable);
nm_clear_g_cancellable(&priv->get_managed_objects_cancellable);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_dbus_connection_signal(priv->dbus_connection, &priv->dbsid_nm_object_manager);
nm_clear_g_dbus_connection_signal(priv->dbus_connection,
&priv->dbsid_dbus_properties_properties_changed);
nm_clear_g_dbus_connection_signal(priv->dbus_connection,
&priv->dbsid_nm_settings_connection_updated);
nm_clear_g_dbus_connection_signal(priv->dbus_connection,
&priv->dbsid_nm_connection_active_state_changed);
nm_clear_g_dbus_connection_signal(priv->dbus_connection,
&priv->dbsid_nm_vpn_connection_state_changed);
nm_clear_g_dbus_connection_signal(priv->dbus_connection, &priv->dbsid_nm_check_permissions);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
if (priv->permissions_state != NM_TERNARY_DEFAULT) {
priv->permissions_state = NM_TERNARY_DEFAULT;
permissions_state_changed = TRUE;
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->permissions) {
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
gs_free guint8 *old_permissions = g_steal_pointer(&priv->permissions);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
_emit_permissions_changed(self, old_permissions, NULL);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
if (permissions_state_changed)
_notify(self, PROP_PERMISSIONS_STATE);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(c_list_is_empty(&priv->obj_changed_lst_head));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
dbus_objects_lst_heads = ((CList *[]) {
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
&priv->dbus_objects_lst_head_on_dbus,
&priv->dbus_objects_lst_head_with_nmobj_not_ready,
&priv->dbus_objects_lst_head_with_nmobj_ready,
NULL,
});
for (i = 0; dbus_objects_lst_heads[i]; i++) {
c_list_for_each_entry (dbobj, dbus_objects_lst_heads[i], dbus_objects_lst) {
NMLDBusObjIfaceData *db_iface_data;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(c_list_is_empty(&dbobj->obj_changed_lst));
c_list_for_each_entry (db_iface_data, &dbobj->iface_lst_head, iface_lst)
db_iface_data->iface_removed = TRUE;
nml_dbus_object_obj_changed_link(self, dbobj, NML_DBUS_OBJ_CHANGED_TYPE_DBUS);
clients/tests: add python test script for nmcli tests Add a test which runs nmcli against our stub NetworkManager service and compares the output. The output formats of nmcli are complicated and not easily understood. For example how --mode tabular|multiline interacts with selecting output-fields (--fields) and output modes ([default]|--terse|--pretty). Also, there are things like `nmcli connection show --order $FIELD_SPEC`. We need unit tests to ensure that we don't change the output accidentally.
2018-05-04 09:02:53 +02:00
}
libnm/udev: cache and reuse udev instance for NMDevice No need to create a separate NMUdevClient instance for all devices. Instead, have one "struct udev" instance in NMClient and pass it down during object construction.
2017-03-20 12:27:17 +01:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_dbus_handle_changes(self, "release-all", FALSE);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/* We require that when we remove all D-Bus interfaces, that all object will go
* away. Note that a NMLDBusObject can be alive due to a NMLDBusObjWatcher, but
* even those should be all cleaned up. */
nm_assert(c_list_is_empty(&priv->obj_changed_lst_head));
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_watched_only));
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_on_dbus));
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_with_nmobj_not_ready));
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_with_nmobj_ready));
libnm: fix re-entrancy of NMClient.dispose() for _init_release_all() GObject's dispose() functions may be called multiple times to break reference cycles. As dispose() calls _init_release_all(), the object might already be partially destroyed. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-12-08 15:41:46 +01:00
nm_assert(nm_g_hash_table_size(priv->dbus_objects) == 0);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
static void
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
name_owner_changed(NMClient *self, const char *name_owner)
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
gboolean changed;
gs_free char *old_name_owner_free = NULL;
const char *old_name_owner;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_auto_pop_gmaincontext GMainContext *dbus_context = NULL;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
name_owner = nm_str_not_empty(name_owner);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
changed = !nm_streq0(priv->name_owner, name_owner);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!name_owner && priv->main_context != priv->dbus_context) {
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
gs_unref_object GObject *old_context_busy_watcher = NULL;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "resync main context as we have no name owner");
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_dbus_connection_signal(priv->dbus_connection, &priv->name_owner_changed_id);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/* Our instance was initialized synchronously. Usually we must henceforth
* stick to a internal main context. But now we have no name-owner...
* at this point, we anyway are going to do a full resync. Swap the main
* contexts again. */
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
old_context_busy_watcher = g_steal_pointer(&priv->context_busy_watcher);
priv->context_busy_watcher = g_object_ref(
g_object_get_qdata(old_context_busy_watcher, nm_context_busy_watcher_quark()));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_main_context_ref(priv->main_context);
g_main_context_unref(priv->dbus_context);
priv->dbus_context = priv->main_context;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_context = nm_g_main_context_push_thread_default_if_necessary(priv->dbus_context);
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/* we need to sync again... */
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_assert_main_context_is_current_thread_default(self, dbus_context);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->name_owner_changed_id =
nm_dbus_connection_signal_subscribe_name_owner_changed(priv->dbus_connection,
NM_DBUS_SERVICE,
name_owner_changed_cb,
self,
NULL);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
name_owner_get_call(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
} else
dbus_context = nm_g_main_context_push_thread_default_if_necessary(priv->dbus_context);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed) {
NML_NMCLIENT_LOG_D(self,
"name owner changed: %s%s%s -> %s%s%s",
NM_PRINT_FMT_QUOTE_STRING(priv->name_owner),
NM_PRINT_FMT_QUOTE_STRING(name_owner));
old_name_owner_free = priv->name_owner;
priv->name_owner = g_strdup(name_owner);
old_name_owner = old_name_owner_free;
} else
old_name_owner = priv->name_owner;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed)
_notify(self, PROP_DBUS_NAME_OWNER);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed && old_name_owner)
_init_release_all(self);
libnm: refactor name-owner check for object-manager
2017-06-21 15:57:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (changed && priv->name_owner)
_init_fetch_all(self);
libnm: refactor name-owner check for object-manager
2017-06-21 15:57:07 +02:00
libnm/client: don't consider client running if we got no NMManager We silently tolerate NetworkManager not responding at all (easily reproduced with e.g. pkill -STOP NetworkManager): $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[23540]: <debug> [3316.81989] nmclient[ddafb84b8deebe4a]: new NMClient instance libnm-dbus[23540]: <debug> [3316.81998] nmclient[ddafb84b8deebe4a]: starting async initialization... libnm-dbus[23540]: <debug> [3316.82461] nmclient[ddafb84b8deebe4a]: name owner changed: (null) -> ":1.2" libnm-dbus[23540]: <debug> [3316.82464] nmclient[ddafb84b8deebe4a]: fetch all libnm-dbus[23540]: <debug> [3341.85715] nmclient[ddafb84b8deebe4a]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[23540]: <debug> [3341.85740] nmclient[ddafb84b8deebe4a]: async init complete with success Error: dummy666 - no such connection profile. libnm-dbus[23540]: <debug> [3341.86723] nmclient[ddafb84b8deebe4a]: release all libnm-dbus[23540]: <debug> [3341.86798] nmclient[ddafb84b8deebe4a]: disposed $ As a comment in _dbus_get_managed_objects_cb() explains, this is sort of intentional. NetworkManager might just be shutting down and the libnm users will eventually see the objects once a new daemon starts up. This may make some sense for long-running clients ("nmcli monitor", various desktop environments), but not for one-shot invocations that require the daemon running, such as those of "nmcli c ...". Let's not consider the client running unless we actually got the manager object. That way the error message will make more sense: $ LIBNM_CLIENT_DEBUG=trace nmcli c show dummy666 libnm-dbus[24730]: <debug> [5360.95480] nmclient[8cb898d3c891e210]: new NMClient instance libnm-dbus[24730]: <debug> [5360.95487] nmclient[8cb898d3c891e210]: starting async initialization... libnm-dbus[24730]: <debug> [5360.95901] nmclient[8cb898d3c891e210]: name owner changed: (null) -> ":1.2" libnm-dbus[24730]: <debug> [5360.95904] nmclient[8cb898d3c891e210]: fetch all libnm-dbus[24730]: <debug> [5385.98487] nmclient[8cb898d3c891e210]: GetManagedObjects() call failed: Timeout was reached libnm-dbus[24730]: <debug> [5385.98497] nmclient[8cb898d3c891e210]: async init complete with success Error: NetworkManager is not running. libnm-dbus[24730]: <debug> [5385.98571] nmclient[8cb898d3c891e210]: release all libnm-dbus[24730]: <debug> [5385.98698] nmclient[8cb898d3c891e210]: disposed $ https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1502
2023-01-16 08:10:37 +01:00
_set_nm_running(self, FALSE);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->init_data) {
nm_auto_pop_gmaincontext GMainContext *main_context = NULL;
libnm: refactor name-owner check for object-manager
2017-06-21 15:57:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->main_context != priv->dbus_context)
main_context = nm_g_main_context_push_thread_default_if_necessary(priv->main_context);
_init_start_check_complete(self);
}
}
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
static void
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
name_owner_changed_cb(GDBusConnection *connection,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *sender_name,
const char *object_path,
const char *interface_name,
const char *signal_name,
GVariant *parameters,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gpointer user_data)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
const char *new_owner;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!g_variant_is_of_type(parameters, G_VARIANT_TYPE("(sss)")))
return;
priv = NM_CLIENT_GET_PRIVATE(self);
if (priv->name_owner_get_cancellable)
libnm: move check for pending_init to init_async_complete()
2017-06-21 15:34:34 +02:00
return;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_variant_get(parameters, "(&s&s&s)", NULL, NULL, &new_owner);
name_owner_changed(self, new_owner);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
static void
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
name_owner_get_cb(GObject *source, GAsyncResult *result, gpointer user_data)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self;
NMClientPrivate *priv;
gs_unref_object GObject *context_busy_watcher = NULL;
gs_unref_variant GVariant *ret = NULL;
gs_free_error GError *error = NULL;
const char *name_owner = NULL;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
nm_utils_user_data_unpack(user_data, &self, &context_busy_watcher);
ret = g_dbus_connection_call_finish(G_DBUS_CONNECTION(source), result, &error);
all: add nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_or_disposing() Most callers would pass FALSE to nm_utils_error_is_cancelled(). That's not very useful. Split the two functions and have nm_utils_error_is_cancelled() and nm_utils_error_is_cancelled_is_disposing().
2020-01-23 10:36:24 +01:00
if (!ret && nm_utils_error_is_cancelled(error))
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv = NM_CLIENT_GET_PRIVATE(self);
libnm: assert in async_inited_obj_nm() for existing pending_init count
2017-06-21 15:20:19 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_clear_object(&priv->name_owner_get_cancellable);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
if (ret)
g_variant_get(ret, "(&s)", &name_owner);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
name_owner_changed(self, name_owner);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
}
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
static void
name_owner_get_call(NMClient *self)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
nm_assert(!priv->name_owner_get_cancellable);
priv->name_owner_get_cancellable = g_cancellable_new();
g_dbus_connection_call(priv->dbus_connection,
DBUS_SERVICE_DBUS,
DBUS_PATH_DBUS,
DBUS_INTERFACE_DBUS,
"GetNameOwner",
g_variant_new("(s)", NM_DBUS_SERVICE),
G_VARIANT_TYPE("(s)"),
G_DBUS_CALL_FLAGS_NONE,
NM_DBUS_DEFAULT_TIMEOUT_MSEC,
priv->name_owner_get_cancellable,
name_owner_get_cb,
nm_utils_user_data_pack(self, g_object_ref(priv->context_busy_watcher)));
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/*****************************************************************************/
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
static inline gboolean
_nml_cleanup_context_busy_watcher_on_idle_cb(gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
nm_auto_unref_gmaincontext GMainContext *context = NULL;
gs_unref_object GObject *context_busy_watcher = NULL;
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
nm_utils_user_data_unpack(user_data, &context, &context_busy_watcher);
nm_assert(context);
nm_assert(G_IS_OBJECT(context_busy_watcher));
return G_SOURCE_REMOVE;
}
void
nml_cleanup_context_busy_watcher_on_idle(GObject *context_busy_watcher_take, GMainContext *context)
{
gs_unref_object GObject *context_busy_watcher = g_steal_pointer(&context_busy_watcher_take);
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GSource *cleanup_source;
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
nm_assert(G_IS_OBJECT(context_busy_watcher));
nm_assert(context);
/* Technically, we cancelled all pending actions (and these actions
* (GTask) keep the context_busy_watcher object alive). Also, we passed
* no destroy notify to g_dbus_connection_signal_subscribe().
* That means, there should be no other unaccounted GSource'es left.
*
libnm: add comment about context-busy-watcher and g_dbus_connection_signal_unsubscribe()
2021-03-29 10:28:39 +02:00
* Another reason is g_dbus_connection_signal_unsubscribe() which does
* not guarantee that everything was unsubscribed right away. Instead,
* there might still be idle actions queued (which will be thrown away
* once processed). Still, that means, the caller must continue to iterate
* the main context afterwards. Maybe we should pass a GDestroyNotify to
* g_dbus_connection_signal_subscribe() which handles an additional reference
* to the context_busy_watcher object. That would be the proper way to handle
* this. We don't do that, so in practice at this point there might still
* be some idle actions (with G_PRIORITY_DEFAULT) pending. As we schedule here
* another idle action with lower priority, we handle those cases without
* using a GDestroyNotify.
*
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
* However, we really need to be sure that the context_busy_watcher's
* lifetime matches the time that the context is busy. That is especially
* important with synchronous initialization, where the context-busy-watcher
* keeps the inner GMainContext integrated in the caller's.
* We must not g_source_destroy() that integration too early.
*
* So to be really sure all this is given, always schedule one last
* cleanup idle action with low priority. This should be the last
* thing related to this instance that keeps the context busy.
*
* Note that we could also *not* take a reference on @context
* and unref @context_busy_watcher via the GDestroyNotify. That would
* allow for the context to be wrapped up early, and when the last user
* gives up the reference to the context, the destroy notify could complete
* without even invoke the idle handler. However, that destroy notify may
* not be called in the right thread. So, we want to be sure that we unref
* the context-busy-watcher in the right context. Hence, we always take an
* additional reference and always cleanup in the idle handler. This means:
* the user *MUST* always keep iterating the context after NMClient got destroyed.
* But that is not a severe limitation, because the user anyway must be prepared
* to do that. That is because in many cases it is necessary anyway (and the user
libnm: add comment about context-busy-watcher and g_dbus_connection_signal_unsubscribe()
2021-03-29 10:28:39 +02:00
* wouldn't know a priory when not). This way, it is just always necessary.
*
* It might be nice to use G_DEFAULT_PRIORITY here to minimize how long the
* instance stays alive. Probably that would be fine, even without passing
* a GDestroyNotify to g_dbus_connection_signal_subscribe(). But to be extra
* careful, use a low priority.
**/
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
cleanup_source =
nm_g_idle_source_new(G_PRIORITY_LOW + 10,
_nml_cleanup_context_busy_watcher_on_idle_cb,
nm_utils_user_data_pack(g_main_context_ref(context),
g_steal_pointer(&context_busy_watcher)),
NULL);
g_source_attach(cleanup_source, context);
g_source_unref(cleanup_source);
}
/*****************************************************************************/
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
static void
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start_complete(NMClient *self, GError *error_take)
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
{
libnm: add "initialized-{good,bad}" flags to NMClientInstanceFlags When using async initialization with GAsyncInitable, the user usually can only know that initialization is complete by passing a callback. In simple cases, that can be cumbersome. Also expose a flag that allows to poll that information. Reuse the existing NM_CLIENT_INSTANCE_FLAGS for that. There is an ugliness here, that suddenly there are instance flags that cannot be set, but are still returned by the getter. But as this is a relatively obscure feature, it seems more lightweight to implement it this way (instead of adding a separate property and getter function).
2022-10-06 10:13:14 +02:00
gs_unref_object NMClient *self_keep_alive = g_object_ref(self);
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(
self,
"%s init complete with %s%s%s",
libnm: factor out nml_init_data_return() for reuse
2019-12-18 17:12:45 +01:00
priv->init_data->is_sync ? "sync" : "async",
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NM_PRINT_FMT_QUOTED(error_take, "error: ", error_take->message, "", "success"));
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: add "initialized-{good,bad}" flags to NMClientInstanceFlags When using async initialization with GAsyncInitable, the user usually can only know that initialization is complete by passing a callback. In simple cases, that can be cumbersome. Also expose a flag that allows to poll that information. Reuse the existing NM_CLIENT_INSTANCE_FLAGS for that. There is an ugliness here, that suddenly there are instance flags that cannot be set, but are still returned by the getter. But as this is a relatively obscure feature, it seems more lightweight to implement it this way (instead of adding a separate property and getter function).
2022-10-06 10:13:14 +02:00
priv->instance_flags |= (error_take ? NM_CLIENT_INSTANCE_FLAGS_INITIALIZED_BAD
: NM_CLIENT_INSTANCE_FLAGS_INITIALIZED_GOOD);
_notify(self, PROP_INSTANCE_FLAGS);
libnm: factor out nml_init_data_return() for reuse
2019-12-18 17:12:45 +01:00
nml_init_data_return(g_steal_pointer(&priv->init_data), error_take);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
}
static void
_init_start_check_complete(NMClient *self)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
_assert_main_context_is_current_thread_default(self, main_context);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!priv->init_data)
return;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->get_managed_objects_cancellable) {
/* still initializing. Wait. */
return;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
#if NM_MORE_ASSERTS > 10
{
NMLDBusObject *dbobj;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_for_each_entry (dbobj,
&priv->dbus_objects_lst_head_with_nmobj_not_ready,
dbus_objects_lst) {
NML_NMCLIENT_LOG_T(self, "init-start waiting for %s", dbobj->dbus_path->str);
break;
}
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
#endif
if (!c_list_is_empty(&priv->dbus_objects_lst_head_with_nmobj_not_ready))
return;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start_complete(self, NULL);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
static void
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start_cancelled_cb(GCancellable *cancellable, gpointer user_data)
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
{
libnm: fix crash during the restart after killed When cloud-init job (metadata service crawler) starts, it sends the SIGTERM signal to nm-cloud-setup and force the nm-cloud-setup to restart, however, because the error is not initialized as NULL in `_init_start_cancelled_cb()` before it is set, nm-cloud-setup will hit a dumped core. TO fix it, initialize the error as NULL in `_init_start_cancelled_cb()`. https://bugzilla.redhat.com/show_bug.cgi?id=2027674 Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') Backtrace: #0 g_logv (log_domain=0x7f833a872071 "GLib", log_level=G_LOG_LEVEL_WARNING, format=<optimized out>, args=<optimized out>) at ../glib/gmessages.c:1413 #1 0x00007f833a81f043 in g_log (log_domain=<optimized out>, log_level=<optimized out>, format=<optimized out>) at ../glib/gmessages.c:1451 #2 0x00007f833ab97230 in nm_utils_error_set_cancelled (is_disposing=<optimized out>, instance_name=<optimized out>, error=0x7ffff79cb980) at src/libnm-glib-aux/nm-shared-utils.c:2599 #3 nm_utils_error_set_cancelled (is_disposing=0, instance_name=0x0, error=0x7ffff79cb980) at src/libnm-glib-aux/nm-shared-utils.c:2590 #4 _init_start_cancelled_cb (cancellable=<optimized out>, user_data=0x5640ca292150) at src/libnm-client-impl/nm-client.c:7324 #5 _init_start_cancelled_cb (cancellable=<optimized out>, user_data=0x5640ca292150) at src/libnm-client-impl/nm-client.c:7307 #6 0x00007f833a93094a in _g_closure_invoke_va (param_types=0x0, n_params=<optimized out>, args=0x7ffff79cbb40, instance=0x5640ca267020, return_value=0x0, closure=0x5640ca29d430) at ../gobject/gclosure.c:873 #7 g_signal_emit_valist (instance=0x5640ca267020, signal_id=<optimized out>, detail=0, var_args=var_args@entry=0x7ffff79cbb40) at ../gobject/gsignal.c:3406 #8 0x00007f833a930a93 in g_signal_emit (instance=instance@entry=0x5640ca267020, signal_id=<optimized out>, detail=detail@entry=0) at ../gobject/gsignal.c:3553 #9 0x00007f833a9a6475 in g_cancellable_cancel (cancellable=0x5640ca267020) at ../gio/gcancellable.c:513 #10 g_cancellable_cancel (cancellable=0x5640ca267020) at ../gio/gcancellable.c:487 #11 0x00005640ca1a8bd4 in sigterm_handler (user_data=0x5640ca267020) at src/nm-cloud-setup/main.c:599 #12 0x00007f833a819d4f in g_main_dispatch (context=0x5640ca268ef0) at ../glib/gmain.c:3337 #13 g_main_context_dispatch (context=0x5640ca268ef0) at ../glib/gmain.c:4055 #14 0x00007f833a86e608 in g_main_context_iterate.constprop.0 (context=0x5640ca268ef0, block=block@entry=1, dispatch=dispatch@entry=1, self=<optimized out>) at ../glib/gmain.c:4131 #15 0x00007f833a819463 in g_main_loop_run (loop=0x5640ca24fdb0) at ../glib/gmain.c:4329 #16 0x00005640ca1a6d04 in nmc_client_new_waitsync (cancellable=0x5640ca267020, out_nmc=0x7ffff79cbfa0, error=0x7ffff79cbf98, first_property_name=0x5640ca1b11db "instance-flags", first_property_name=0x5640ca1b11db "instance-flags") at src/libnm-client-aux-extern/nm-libnm-aux.c:129 #17 0x00005640ca1a3863 in main (argc=1, argv=<optimized out>) at src/nm-cloud-setup/main.c:639
2022-03-08 15:58:38 -05:00
NMClient *self = user_data;
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
GError *error = NULL;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
libnm: cleanup handling of cancellation for initializing NMClient
2021-03-29 09:06:50 +02:00
nm_assert(priv->init_data);
nm_assert(priv->init_data->cancellable == cancellable);
if (priv->init_data->cancelled_id == 0) {
/* this only can happen if the cancellable was already cancelled initially.
* In this case we are called synchronously. Do nothing, and let the caller
* handle it. */
return;
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_utils_error_set_cancelled(&error, FALSE, NULL);
_init_start_complete(self, error);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static gboolean
_init_start_cancel_on_idle_cb(gpointer user_data)
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
{
libnm: fix uninitialized GError in NMClient's _init_start_cancel_on_idle_cb() Found with `git grep 'GError.*[^,)];'| grep ' *= *NULL;' -v` Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2022-03-09 23:00:21 +01:00
NMClient *self = user_data;
GError *error = NULL;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_utils_error_set_cancelled(&error, FALSE, NULL);
_init_start_complete(self, error);
return G_SOURCE_CONTINUE;
}
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
static void
_init_start_with_bus(NMClient *self)
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->init_data->cancellable) {
libnm: cleanup handling of cancellation for initializing NMClient
2021-03-29 09:06:50 +02:00
gulong id;
id = g_cancellable_connect(priv->init_data->cancellable,
G_CALLBACK(_init_start_cancelled_cb),
self,
NULL);
if (id == 0) {
priv->init_data->cancel_on_idle_source =
all: change scheduling priority for idle actions to G_PRIORITY_DEFAULT_IDLE g_idle_add() uses G_PRIORITY_DEFAULT_IDLE priority. Most of the time we don't care much about the priority. But at the places that this patch changes, I think that using G_PRIORITY_DEFAULT_IDLE (and following g_idle_add()) is more correct. The reason for this is not very strong, except that it's probably the better choice. And the old choice was made because I didn't realize that g_idle_add() uses another default priority. Hence, the old choice was not for good reasons either.
2022-03-08 14:29:51 +01:00
nm_g_idle_source_new(G_PRIORITY_DEFAULT_IDLE,
_init_start_cancel_on_idle_cb,
self,
NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_source_attach(priv->init_data->cancel_on_idle_source, priv->main_context);
Revert "libnm: don't use async constructor for GDBusObjectManager" Strangely, this breaks systemctl restart NetworkManager nmcli connection up "$NAME" It seems that with this change, libnm misses some events from D-Bus. It looks like there is something seriously broken. Before fixing it, revert the previous state. https://bugzilla.redhat.com/show_bug.cgi?id=1450075 This reverts commit 529d620a59f42e3f06bd4e7de5c817f175803c0d.
2017-06-28 18:42:04 +02:00
return;
libnm: fix leaking init_data in got_object_manager() Only happens if there are no objects, which would be very unusual.
2017-06-21 15:32:23 +02:00
}
libnm: cleanup handling of cancellation for initializing NMClient
2021-03-29 09:06:50 +02:00
priv->init_data->cancelled_id = id;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_assert_main_context_is_current_thread_default(self, dbus_context);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv->name_owner_changed_id =
nm_dbus_connection_signal_subscribe_name_owner_changed(priv->dbus_connection,
NM_DBUS_SERVICE,
name_owner_changed_cb,
self,
NULL);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
name_owner_get_call(self);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
static void
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start_bus_get_cb(GObject *source, GAsyncResult *result, gpointer user_data)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = user_data;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv;
GDBusConnection *dbus_connection;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GError *error = NULL;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(NM_IS_CLIENT(self));
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_connection = g_bus_get_finish(result, &error);
libnm: add NM_CLIENT_DBUS_CONNECTION property The used GDBusConnection should be configurable when creating the NMClient instance. Automatically choosing one of the g_bus_get() singletons is fine by default, but it's an unnecessary limitation.
2019-10-12 14:59:23 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!dbus_connection) {
_init_start_complete(self, error);
return;
}
priv = NM_CLIENT_GET_PRIVATE(self);
priv->dbus_connection = dbus_connection;
_init_start_with_bus(self);
libnm: emit property changed signal when setting NM_CLIENT_DBUS_CONNECTION
2019-12-18 17:13:06 +01:00
_notify(self, PROP_DBUS_CONNECTION);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
static void
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start(NMClient *self)
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
{
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self,
"starting %s initialization...",
priv->init_data->is_sync ? "sync" : "async");
libnm: clear NMClientPrivate.object_manager in name_owner_changed() Don't leave dangling pointers.
2017-06-20 14:13:16 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (!priv->dbus_connection) {
g_bus_get(_nm_dbus_bus_type(), priv->init_data->cancellable, _init_start_bus_get_cb, self);
return;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start_with_bus(self);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
}
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
/*****************************************************************************/
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
static void
libnm: split "NMManager" out of NMClient Clone NMClient as NMManager, and make NMClient just be a wrapper around that new class.
2014-09-24 13:14:48 -04:00
get_property(GObject *object, guint prop_id, GValue *value, GParamSpec *pspec)
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = NM_CLIENT(object);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(object);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
switch (prop_id) {
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
case PROP_INSTANCE_FLAGS:
g_value_set_uint(value, priv->instance_flags);
break;
libnm: add NM_CLIENT_DBUS_CONNECTION property The used GDBusConnection should be configurable when creating the NMClient instance. Automatically choosing one of the g_bus_get() singletons is fine by default, but it's an unnecessary limitation.
2019-10-12 14:59:23 +02:00
case PROP_DBUS_CONNECTION:
g_value_set_object(value, priv->dbus_connection);
break;
libnm: add NM_CLIENT_DBUS_NAME_OWNER property It's not yet implemented. But obviously it's interesting to get the name owner to which the NMClient is currently connected. Note only that: the name-owner property really says whether NM is currently running or not.
2019-10-15 16:08:36 +02:00
case PROP_DBUS_NAME_OWNER:
g_value_set_string(value, nm_client_get_dbus_name_owner(self));
break;
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
case PROP_NM_RUNNING:
g_value_set_boolean(value, nm_client_get_nm_running(self));
break;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
/* Manager properties. */
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_VERSION:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_string(value, nm_client_get_version(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_STATE:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_enum(value, nm_client_get_state(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_STARTUP:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_boolean(value, nm_client_get_startup(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_NETWORKING_ENABLED:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_boolean(value, nm_client_networking_get_enabled(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_WIRELESS_ENABLED:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_boolean(value, nm_client_wireless_get_enabled(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_WIRELESS_HARDWARE_ENABLED:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_set_boolean(value, nm_client_wireless_hardware_get_enabled(self));
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
break;
libnm: support radio flags
2022-03-21 10:20:11 +01:00
case PROP_RADIO_FLAGS:
g_value_set_uint(value, priv->nm.radio_flags);
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_WWAN_ENABLED:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_boolean(value, nm_client_wwan_get_enabled(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_WWAN_HARDWARE_ENABLED:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_set_boolean(value, nm_client_wwan_hardware_get_enabled(self));
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_WIMAX_ENABLED:
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
g_value_set_boolean(value, FALSE);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_WIMAX_HARDWARE_ENABLED:
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
g_value_set_boolean(value, FALSE);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_ACTIVE_CONNECTIONS:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_take_boxed(value,
_nm_utils_copy_object_array(nm_client_get_active_connections(self)));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_CONNECTIVITY:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_enum(value, nm_client_get_connectivity(self));
break;
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
case PROP_CONNECTIVITY_CHECK_AVAILABLE:
g_value_set_boolean(value, nm_client_connectivity_check_get_available(self));
break;
case PROP_CONNECTIVITY_CHECK_ENABLED:
g_value_set_boolean(value, nm_client_connectivity_check_get_enabled(self));
break;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
case PROP_CONNECTIVITY_CHECK_URI:
g_value_set_string(value, nm_client_connectivity_check_get_uri(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_PRIMARY_CONNECTION:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_object(value, nm_client_get_primary_connection(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_ACTIVATING_CONNECTION:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_set_object(value, nm_client_get_activating_connection(self));
break;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
case PROP_DEVICES:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_take_boxed(value, _nm_utils_copy_object_array(nm_client_get_devices(self)));
break;
nm-manager: add 'metered' property This introduces a global metered property which makes easier for clients to obtain the metered status of the current primary connection.
2015-06-03 09:15:24 +02:00
case PROP_METERED:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_set_uint(value, nm_client_get_metered(self));
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
break;
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
case PROP_ALL_DEVICES:
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
g_value_take_boxed(value, _nm_utils_copy_object_array(nm_client_get_all_devices(self)));
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
break;
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
case PROP_CHECKPOINTS:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_take_boxed(value, _nm_utils_copy_object_array(nm_client_get_checkpoints(self)));
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
break;
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
case PROP_VERSION_INFO:
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
case PROP_CAPABILITIES:
{
const guint32 *arr;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GArray *out;
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
gsize len;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
arr = (prop_id == PROP_VERSION_INFO) ? nm_client_get_version_info(self, &len)
: nm_client_get_capabilities(self, &len);
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
if (arr) {
out = g_array_new(TRUE, FALSE, sizeof(guint32));
g_array_append_vals(out, arr, len);
} else
out = NULL;
g_value_take_boxed(value, out);
} break;
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
case PROP_PERMISSIONS_STATE:
g_value_set_enum(value, priv->permissions_state);
break;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
/* Settings properties. */
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
case PROP_CONNECTIONS:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_take_boxed(value, _nm_utils_copy_object_array(nm_client_get_connections(self)));
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
break;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
case PROP_HOSTNAME:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_set_string(value, priv->settings.hostname);
libnm: use the o.fd.DBus.ObjectManager API for object management This speeds up the initial object tree load significantly. Also, it reduces the object management complexity by shifting the duties to GDBusObjectManager. The lifetime of all NMObjects is now managed by the NMClient via the object manager. The NMClient creates the NMObjects for GDBus objects, triggers the initialization and serves as an object registry (replaces the nm-cache). The ObjectManager uses the o.fd.DBus.ObjectManager API to learn of the object creation, removal and property changes. It takes care of the property changes so that we don't have to and lets us always see a consistent object state. Thus at the time we learn of a new object we already know its properties. The NMObject unfortunately can't be made synchronously initializable as the NMRemoteConnection's settings are not managed with standard o.fd.DBus Properties and ObjectManager APIs and thus are not known to the ObjectManager. Thus most of the asynchronous object property changing code in nm-object.c is preserved. The objects notify the properties that reference them of their initialization in from their init_finish() methods, thus the asynchronously created objects are not allowed to fail creation (or the dependees would wait forever). Not a problem -- if a connection can't get its Settings, it's either invisible or being removed (presumably we'd learn of the removal from the object manager soon). The NMObjects can't be created by the object manager itself, since we can't determine the resulting object type in proxy_type() yet (we can't tell from the name and can't access the interface list). Therefore the GDBusObject is coupled with a NMObject later on. Lastly, now that all the objects are managed by the object manager, the NMRemoteSettings and NMManager go away when the daemon is stopped. The complexity of dealing with calls to NMClient that would require any of the resources that these objects manage (connection or device lists, etc.) had to be moved to NMClient. The bright side is that his allows for removal all of the daemon presence tracking from NMObject.
2016-10-18 16:35:07 +02:00
break;
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
case PROP_CAN_MODIFY:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_set_boolean(value, priv->settings.can_modify);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
break;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
/* DNS properties */
case PROP_DNS_MODE:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_set_string(value, nm_client_get_dns_mode(self));
break;
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
case PROP_DNS_RC_MANAGER:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_set_string(value, nm_client_get_dns_rc_manager(self));
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
break;
case PROP_DNS_CONFIGURATION:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_value_take_boxed(value,
_nm_utils_copy_array(nm_client_get_dns_configuration(self),
(NMUtilsCopyFunc) nm_dns_entry_dup,
(GDestroyNotify) nm_dns_entry_unref));
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
break;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
default:
G_OBJECT_WARN_INVALID_PROPERTY_ID(object, prop_id, pspec);
break;
}
}
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
static void
set_property(GObject *object, guint prop_id, const GValue *value, GParamSpec *pspec)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = NM_CLIENT(object);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
gboolean b;
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
guint v_uint;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
switch (prop_id) {
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
case PROP_INSTANCE_FLAGS:
/* construct */
v_uint = g_value_get_uint(value);
libnm: add "initialized-{good,bad}" flags to NMClientInstanceFlags When using async initialization with GAsyncInitable, the user usually can only know that initialization is complete by passing a callback. In simple cases, that can be cumbersome. Also expose a flag that allows to poll that information. Reuse the existing NM_CLIENT_INSTANCE_FLAGS for that. There is an ugliness here, that suddenly there are instance flags that cannot be set, but are still returned by the getter. But as this is a relatively obscure feature, it seems more lightweight to implement it this way (instead of adding a separate property and getter function).
2022-10-06 10:13:14 +02:00
/* Silently ignore "initialized-{good,bad}" flags. They are only set internally
* and cannot be change by the user. However, accept the caller to set them,
* so that
* nmc.props.instance_flags = nmc.props.instance_flags | NM.ClientInstanceFlags.NO_AUTO_FETCH_PERMISSIONS
* works. */
v_uint &= ~((guint) (NM_CLIENT_INSTANCE_FLAGS_INITIALIZED_GOOD
| NM_CLIENT_INSTANCE_FLAGS_INITIALIZED_BAD));
g_return_if_fail(!NM_FLAGS_ANY(v_uint, ~((guint) NM_CLIENT_INSTANCE_FLAGS_ALL_WRITABLE)));
v_uint &= ((guint) NM_CLIENT_INSTANCE_FLAGS_ALL_WRITABLE);
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
if (!priv->instance_flags_constructed) {
priv->instance_flags_constructed = TRUE;
priv->instance_flags = v_uint;
nm_assert((guint) priv->instance_flags == v_uint);
} else {
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
NMClientInstanceFlags flags = v_uint;
typo fix: allow to -> allow one to Detected by lintian: Example: I: network-manager: typo-in-manual-page "allow to" "allow one to" [usr/share/man/man5/NetworkManager.conf.5.gz:1392]
2025-02-28 18:37:15 +01:00
/* After object construction, we only allow one to toggle certain flags and
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
* ignore all other flags. */
if ((priv->instance_flags ^ flags)
& NM_CLIENT_INSTANCE_FLAGS_NO_AUTO_FETCH_PERMISSIONS) {
if (NM_FLAGS_HAS(flags, NM_CLIENT_INSTANCE_FLAGS_NO_AUTO_FETCH_PERMISSIONS))
priv->instance_flags |= NM_CLIENT_INSTANCE_FLAGS_NO_AUTO_FETCH_PERMISSIONS;
else
priv->instance_flags &= ~NM_CLIENT_INSTANCE_FLAGS_NO_AUTO_FETCH_PERMISSIONS;
if (priv->dbsid_nm_check_permissions != 0)
_dbus_check_permissions_start(self);
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
}
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
break;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add NM_CLIENT_DBUS_CONNECTION property The used GDBusConnection should be configurable when creating the NMClient instance. Automatically choosing one of the g_bus_get() singletons is fine by default, but it's an unnecessary limitation.
2019-10-12 14:59:23 +02:00
case PROP_DBUS_CONNECTION:
/* construct-only */
priv->dbus_connection = g_value_dup_object(value);
break;
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
case PROP_NETWORKING_ENABLED:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
b = g_value_get_boolean(value);
if (priv->nm.networking_enabled != b) {
nm_client_networking_set_enabled(self, b, NULL);
/* Let the property value flip when we get the change signal from NM */
}
break;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
case PROP_WIRELESS_ENABLED:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
b = g_value_get_boolean(value);
if (priv->nm.wireless_enabled != b) {
nm_client_wireless_set_enabled(self, b);
/* Let the property value flip when we get the change signal from NM */
}
break;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
case PROP_WWAN_ENABLED:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
b = g_value_get_boolean(value);
if (priv->nm.wwan_enabled != b) {
nm_client_wwan_set_enabled(self, b);
/* Let the property value flip when we get the change signal from NM */
}
break;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
case PROP_CONNECTIVITY_CHECK_ENABLED:
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
b = g_value_get_boolean(value);
if (priv->nm.connectivity_check_enabled != b) {
nm_client_connectivity_check_set_enabled(self, b);
/* Let the property value flip when we get the change signal from NM */
}
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
break;
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
case PROP_WIMAX_ENABLED:
break;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
default:
G_OBJECT_WARN_INVALID_PROPERTY_ID(object, prop_id, pspec);
break;
}
}
/*****************************************************************************/
static gboolean
init_sync(GInitable *initable, GCancellable *cancellable, GError **error)
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
gs_unref_object NMClient *self = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
GMainContext *dbus_context;
GError *local_error = NULL;
GMainLoop *main_loop;
GObject *parent_context_busy_watcher;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_val_if_fail(NM_IS_CLIENT(initable), FALSE);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
self = g_object_ref(NM_CLIENT(initable)); /* keep instance alive. */
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
priv = NM_CLIENT_GET_PRIVATE(self);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_val_if_fail(!priv->dbus_context, FALSE);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/* when using init_sync(), we use a separate internal GMainContext for
* all D-Bus operations and use our regular async-init code. That means,
* also in sync-init, we don't actually block waiting for our D-Bus requests,
* instead, we only block (g_main_loop_run()) for the overall result.
*
* Doing this has a performance overhead. Also, we cannot ever fall back
* to the regular main-context (not unless we lose the main-owner and
* need to re-initialize). The reason is that we receive events on our
* dbus_context, and this cannot be brought in sync -- short of full
* reinitalizing. Therefor, using sync init not only is slower during
* construction of the object, but NMClient will stick to the dual GMainContext
* mode.
*
* Aside from this downside, the solution is good:
*
* - we don't duplicate the implementation of async-init.
* - we don't iterate the main-context of the caller while waiting for
* initialization to happen
* - we still invoke all changes under the main_context of the caller.
* - all D-Bus events strictly go through dbus_context and are in order.
*/
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
dbus_context = g_main_context_new();
priv->dbus_context = g_main_context_ref(dbus_context);
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
/* We have an inner context. Note that if we loose the name owner, we have a chance
* to resync and drop the inner context. That means, requests made against the inner
* context have a different lifetime. Hence, we create a separate tracking
* object. This "wraps" the outer context-busy-watcher and references it, so
libnm: expose nm_context_busy_watcher_integrate_source() as internal API for reuse This will also be useful for NMSecretAgentOld. The mechanics how NMClient handles the GMainContext and the context-busy-watcher apply really to every GObject that uses GDBusConnection and registers to signals. At least, as long as the API provides no shutdown/stop method, because that means shutdown/stop happens when unreferencing the instance, at which point pending operations get cancelled (but they cannot complete right away due to the nature of GTask and g_dbus_connection_call()). If there is a shutdown/stop API, then all pending operations could keep the instance alive, and the instance would sticks around (and keeps the GMainContext busy) until shutdown is completed. Basically, then the instance could be the context-busy-watcher itself. But in existing API which does not require the user to explicitly shutdown, that is not a feasible (backward compatible) addition. But the context-busy-watcher object is.
2019-12-17 13:59:15 +01:00
* that the work together. Grep for nm_context_busy_watcher_quark() to
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
* see how this works. */
parent_context_busy_watcher = g_steal_pointer(&priv->context_busy_watcher);
priv->context_busy_watcher = g_object_new(G_TYPE_OBJECT, NULL);
g_object_set_qdata_full(priv->context_busy_watcher,
libnm: expose nm_context_busy_watcher_integrate_source() as internal API for reuse This will also be useful for NMSecretAgentOld. The mechanics how NMClient handles the GMainContext and the context-busy-watcher apply really to every GObject that uses GDBusConnection and registers to signals. At least, as long as the API provides no shutdown/stop method, because that means shutdown/stop happens when unreferencing the instance, at which point pending operations get cancelled (but they cannot complete right away due to the nature of GTask and g_dbus_connection_call()). If there is a shutdown/stop API, then all pending operations could keep the instance alive, and the instance would sticks around (and keeps the GMainContext busy) until shutdown is completed. Basically, then the instance could be the context-busy-watcher itself. But in existing API which does not require the user to explicitly shutdown, that is not a feasible (backward compatible) addition. But the context-busy-watcher object is.
2019-12-17 13:59:15 +01:00
nm_context_busy_watcher_quark(),
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
parent_context_busy_watcher,
g_object_unref);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_main_context_push_thread_default(dbus_context);
main_loop = g_main_loop_new(dbus_context, FALSE);
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
priv->init_data = nml_init_data_new_sync(cancellable, main_loop, &local_error);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start(self);
g_main_loop_run(main_loop);
g_main_loop_unref(main_loop);
g_main_context_pop_thread_default(dbus_context);
if (priv->main_context != priv->dbus_context) {
libnm: expose nm_context_busy_watcher_integrate_source() as internal API for reuse This will also be useful for NMSecretAgentOld. The mechanics how NMClient handles the GMainContext and the context-busy-watcher apply really to every GObject that uses GDBusConnection and registers to signals. At least, as long as the API provides no shutdown/stop method, because that means shutdown/stop happens when unreferencing the instance, at which point pending operations get cancelled (but they cannot complete right away due to the nature of GTask and g_dbus_connection_call()). If there is a shutdown/stop API, then all pending operations could keep the instance alive, and the instance would sticks around (and keeps the GMainContext busy) until shutdown is completed. Basically, then the instance could be the context-busy-watcher itself. But in existing API which does not require the user to explicitly shutdown, that is not a feasible (backward compatible) addition. But the context-busy-watcher object is.
2019-12-17 13:59:15 +01:00
nm_context_busy_watcher_integrate_source(priv->main_context,
priv->dbus_context,
priv->context_busy_watcher);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_main_context_unref(dbus_context);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (local_error) {
g_propagate_error(error, local_error);
return FALSE;
}
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
return TRUE;
}
/*****************************************************************************/
static void
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
init_async(GAsyncInitable *initable,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
int io_priority,
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GCancellable *cancellable,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
GAsyncReadyCallback callback,
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
gpointer user_data)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClientPrivate *priv;
NMClient *self;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_auto_pop_gmaincontext GMainContext *context = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GTask *task;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_if_fail(NM_IS_CLIENT(initable));
self = NM_CLIENT(initable);
priv = NM_CLIENT_GET_PRIVATE(self);
g_return_if_fail(!priv->dbus_context);
priv->dbus_context = g_main_context_ref(priv->main_context);
context = nm_g_main_context_push_thread_default_if_necessary(priv->main_context);
task = nm_g_task_new(self, cancellable, init_async, callback, user_data);
g_task_set_priority(task, io_priority);
libnm: move InitData to nm-libnm-utils.h for sharing It can be reused.
2019-12-17 14:38:29 +01:00
priv->init_data = nml_init_data_new_async(cancellable, g_steal_pointer(&task));
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_start(self);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
}
static gboolean
init_finish(GAsyncInitable *initable, GAsyncResult *result, GError **error)
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
g_return_val_if_fail(NM_IS_CLIENT(initable), FALSE);
g_return_val_if_fail(nm_g_task_is_valid(result, initable, init_async), FALSE);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
return g_task_propagate_boolean(G_TASK(result), error);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
}
/*****************************************************************************/
static void
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
nm_client_init(NMClient *self)
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
{
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
priv->permissions_state = NM_TERNARY_DEFAULT;
libnm: fix leaking internal GMainContext for synchronously initialized NMClient NMClient makes asynchronous D-Bus calls via g_dbus_connection_call(). This references the current GMainContext to later invoke the asynchronous callback. Even when we cancel the asynchronous call, the callback will still be invoked (later) to complete the request. In particular this means when we destroy (unref) an NMClient, there are quite possibly pending requests in the GMainContext. Although they are cancelled, they keep the GMainContext alive. With synchronous initialization, we have an internal GMainContext. When we destroy the NMClient, we cannot just unhook the integrated source, instead, we need to keep it integrated in the caller's main context, as long as there are pending requests. Add a mechanism to track those pending requests and fix the leak for the internal GMainContext. Also expose the same mechanism to the user via a new API called nm_client_get_context_busy_watcher(). This allows the user to know when it can stop iterating the main context and when all resources are reclaimed. For example the following will lead to a crash: for i in range(1,2000): nmc = NM.Client.new(None) This creates a number of NMClient instances and destroys them again. Note that here the GMainContext is never iterated, because synchronous initialization does not iterate the caller's context. So, while we correctly unref and dispose the created NMClient instances, there are pending requests left in the inner GMainContext. These pile up and soon the program will crash because it runs out of file descriptors. We can have a similar problem with asynchronous initialization, when we create a new GMainContext per client, and don't iterate it after we are done with the client. Note that this patch does not avoid the problem in general. The problem cannot be avoided, the user must iterate the main contex at some point. Otherwise resources (memory and file descriptors) will be leaked. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient') https://gitlab.freedesktop.org/NetworkManager/NetworkManager/merge_requests/347
2019-11-26 00:23:41 +01:00
priv->context_busy_watcher = g_object_new(G_TYPE_OBJECT, NULL);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
c_list_init(&self->obj_base.queue_notify_lst);
c_list_init(&priv->queue_notify_lst_head);
c_list_init(&priv->notify_event_lst_head);
priv->dbus_objects = g_hash_table_new(nm_pdirect_hash, nm_pdirect_equal);
c_list_init(&priv->dbus_objects_lst_head_watched_only);
c_list_init(&priv->dbus_objects_lst_head_on_dbus);
c_list_init(&priv->dbus_objects_lst_head_with_nmobj_not_ready);
c_list_init(&priv->dbus_objects_lst_head_with_nmobj_ready);
c_list_init(&priv->obj_changed_lst_head);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
}
/**
* nm_client_new:
* @cancellable: a #GCancellable, or %NULL
* @error: location for a #GError, or %NULL
*
libnm: better explain nm_client_new()/nm_client_new_async()
2021-03-29 09:10:43 +02:00
* Creates a new #NMClient synchronously.
*
* Note that this will block until a NMClient instance is fully initialized.
* This does nothing beside calling g_initable_new(). You are free to call
* g_initable_new() or g_object_new()/g_initable_init() directly for more
* control, to set GObject properties or get access to the NMClient instance
* while it is still initializing.
*
* Using the synchronous initialization creates an #NMClient instance
libnm/doc: clarify GMainContext handling in NMClient documentation https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/819
2021-04-18 08:34:27 +02:00
* that uses an internal #GMainContext. This context is invisible to the
* user. This introduces an additional overhead that is payed not
* only during object initialization, but for the entire lifetime of
* this object.
* Also, due to this internal #GMainContext, the events are no longer
* in sync with other messages from #GDBusConnection (but all events
* of the NMClient will themselves still be ordered).
* For a serious program, you should therefore avoid these problems by
* using g_async_initable_init_async() or nm_client_new_async() instead.
* The sync initialization is still useful for simple scripts or interactive
* testing for example via pygobject.
libnm: better explain nm_client_new()/nm_client_new_async()
2021-03-29 09:10:43 +02:00
*
* Creating an #NMClient instance can only fail for two reasons. First, if you didn't
* provide a %NM_CLIENT_DBUS_CONNECTION and the call to g_bus_get()
* fails. You can avoid that by using g_initable_new() directly and
* set a D-Bus connection.
* Second, if you cancelled the creation. If you do that, then note
* that after the failure there might still be idle actions pending
* which keep nm_client_get_main_context() alive. That means,
* in that case you must continue iterating the context to avoid
* leaks. See nm_client_get_context_busy_watcher().
*
* Creating an #NMClient instance when NetworkManager is not running
* does not cause a failure.
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
*
* Returns: a new #NMClient or NULL on an error
**/
NMClient *
nm_client_new(GCancellable *cancellable, GError **error)
{
return g_initable_new(NM_TYPE_CLIENT, cancellable, error, NULL);
}
/**
* nm_client_new_async:
* @cancellable: a #GCancellable, or %NULL
* @callback: callback to call when the client is created
* @user_data: data for @callback
*
libnm: better explain nm_client_new()/nm_client_new_async()
2021-03-29 09:10:43 +02:00
* Creates a new #NMClient asynchronously.
* @callback will be called when it is done. Use
* nm_client_new_finish() to get the result.
*
* This does nothing beside calling g_async_initable_new_async(). You are free to
* call g_async_initable_new_async() or g_object_new()/g_async_initable_init_async()
* directly for more control, to set GObject properties or get access to the NMClient
* instance while it is still initializing.
*
* Creating an #NMClient instance can only fail for two reasons. First, if you didn't
* provide a %NM_CLIENT_DBUS_CONNECTION and the call to g_bus_get()
libnm/doc: clarify GMainContext handling in NMClient documentation https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/819
2021-04-18 08:34:27 +02:00
* fails. You can avoid that by using g_async_initable_new_async() directly and
libnm: better explain nm_client_new()/nm_client_new_async()
2021-03-29 09:10:43 +02:00
* set a D-Bus connection.
* Second, if you cancelled the creation. If you do that, then note
* that after the failure there might still be idle actions pending
* which keep nm_client_get_main_context() alive. That means,
* in that case you must continue iterating the context to avoid
* leaks. See nm_client_get_context_busy_watcher().
*
* Creating an #NMClient instance when NetworkManager is not running
* does not cause a failure.
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
**/
void
nm_client_new_async(GCancellable *cancellable, GAsyncReadyCallback callback, gpointer user_data)
{
libnm: don't use GSimpleAsyncResult for nm_client_new_async() As we don't have any data or our own, we don't need a GSimpleAsyncResult/GTask. Just pass the caller's @callback to g_async_initable_new_async().
2019-10-14 13:16:49 +02:00
g_async_initable_new_async(NM_TYPE_CLIENT,
G_PRIORITY_DEFAULT,
cancellable,
callback,
user_data,
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
NULL);
}
/**
* nm_client_new_finish:
* @result: a #GAsyncResult
* @error: location for a #GError, or %NULL
*
* Gets the result of an nm_client_new_async() call.
*
* Returns: a new #NMClient, or %NULL on error
**/
NMClient *
nm_client_new_finish(GAsyncResult *result, GError **error)
{
libnm: don't use GSimpleAsyncResult for nm_client_new_async() As we don't have any data or our own, we don't need a GSimpleAsyncResult/GTask. Just pass the caller's @callback to g_async_initable_new_async().
2019-10-14 13:16:49 +02:00
gs_unref_object GObject *source_object = NULL;
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
GObject *object;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: don't use GSimpleAsyncResult for nm_client_new_async() As we don't have any data or our own, we don't need a GSimpleAsyncResult/GTask. Just pass the caller's @callback to g_async_initable_new_async().
2019-10-14 13:16:49 +02:00
source_object = g_async_result_get_source_object(result);
g_return_val_if_fail(source_object, NULL);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: don't use GSimpleAsyncResult for nm_client_new_async() As we don't have any data or our own, we don't need a GSimpleAsyncResult/GTask. Just pass the caller's @callback to g_async_initable_new_async().
2019-10-14 13:16:49 +02:00
object = g_async_initable_new_finish(G_ASYNC_INITABLE(source_object), result, error);
g_return_val_if_fail(!object || NM_IS_CLIENT(object), FALSE);
return NM_CLIENT(object);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
}
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
static void
constructed(GObject *object)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = NM_CLIENT(object);
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
priv->main_context = g_main_context_ref_thread_default();
G_OBJECT_CLASS(nm_client_parent_class)->constructed(object);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "new NMClient instance");
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
}
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
static void
dispose(GObject *object)
{
format: reformat source tree with clang-format 13.0 We use clang-format for automatic formatting of our source files. Since clang-format is actively maintained software, the actual formatting depends on the used version of clang-format. That is unfortunate and painful, but really unavoidable unless clang-format would be strictly bug-compatible. So the version that we must use is from the current Fedora release, which is also tested by our gitlab-ci. Previously, we were using Fedora 34 with clang-tools-extra-12.0.1-1.fc34.x86_64. As Fedora 35 comes along, we need to update our formatting as Fedora 35 comes with version "13.0.0~rc1-1.fc35". An alternative would be to freeze on version 12, but that has different problems (like, it's cumbersome to rebuild clang 12 on Fedora 35 and it would be cumbersome for our developers which are on Fedora 35 to use a clang that they cannot easily install). The (differently painful) solution is to reformat from time to time, as we switch to a new Fedora (and thus clang) version. Usually we would expect that such a reformatting brings minor changes. But this time, the changes are huge. That is mentioned in the release notes [1] as Makes PointerAligment: Right working with AlignConsecutiveDeclarations. (Fixes https://llvm.org/PR27353) [1] https://releases.llvm.org/13.0.0/tools/clang/docs/ReleaseNotes.html#clang-format
2021-11-09 13:28:54 +01:00
NMClient *self = NM_CLIENT(object);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClientPrivate *priv = NM_CLIENT_GET_PRIVATE(self);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(!priv->init_data);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
self->obj_base.is_disposing = TRUE;
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_cancellable(&priv->name_owner_get_cancellable);
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_dbus_connection_signal(priv->dbus_connection, &priv->name_owner_changed_id);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_free(&priv->name_owner);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_init_release_all(self);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_watched_only));
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_on_dbus));
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_with_nmobj_not_ready));
nm_assert(c_list_is_empty(&priv->dbus_objects_lst_head_with_nmobj_ready));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_assert(c_list_is_empty(&priv->queue_notify_lst_head));
nm_assert(c_list_is_empty(&priv->notify_event_lst_head));
nm_assert(c_list_is_empty(&self->obj_base.queue_notify_lst));
libnm: fix re-entrancy of NMClient.dispose() for _init_release_all() GObject's dispose() functions may be called multiple times to break reference cycles. As dispose() calls _init_release_all(), the object might already be partially destroyed. Fixes: ce0e898fb476 ('libnm: refactor caching of D-Bus objects in NMClient')
2020-12-08 15:41:46 +01:00
nm_assert(nm_g_hash_table_size(priv->dbus_objects) == 0);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nml_dbus_property_o_clear_many(priv->nm.property_o, G_N_ELEMENTS(priv->nm.property_o), NULL);
nml_dbus_property_ao_clear_many(priv->nm.property_ao, G_N_ELEMENTS(priv->nm.property_ao), NULL);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_free(&priv->nm.connectivity_check_uri);
nm_clear_g_free(&priv->nm.version);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nml_dbus_property_ao_clear(&priv->settings.connections, NULL);
nm_clear_g_free(&priv->settings.hostname);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_pointer(&priv->dns_manager.configuration, g_ptr_array_unref);
nm_clear_g_free(&priv->dns_manager.mode);
nm_clear_g_free(&priv->dns_manager.rc_manager);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_pointer(&priv->dbus_objects, g_hash_table_destroy);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
G_OBJECT_CLASS(nm_client_parent_class)->dispose(object);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_pointer(&priv->udev, udev_unref);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
if (priv->context_busy_watcher && priv->dbus_context) {
nml_cleanup_context_busy_watcher_on_idle(g_steal_pointer(&priv->context_busy_watcher),
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
priv->dbus_context);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm/trivial: move code in "nm-client.c"
2019-10-11 11:58:09 +02:00
nm_clear_pointer(&priv->dbus_context, g_main_context_unref);
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
nm_clear_pointer(&priv->main_context, g_main_context_unref);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: keep context-busy-watcher of NMClient alive for one more idle round The context-busy-watch has two purposes: 1) it allows the user to watch whether the NMClient still has pending GSource'es attached to the GMainContext. 2) thereby, it also keeps the inner GMainContext integrated into the caller's (in case of synchronous initialization of NMClient). Especially for 2), we must not get this wrong. Otherwise, we might un-integrate the inner GMainContext too early and it will be leaked indefinitely (because the user has no means to access or iterate it). To be extra careful, extend the lifetime of the context-busy-watcher for one more idle invocation. Theoretically, this should not be necessary, but it's not clear whether something else is still pending. The downside of that extra safety is that it is probably unnecessary in practice. And in case where it is necessary, it hides an actual issue, making it harder to notice and fix it.
2020-01-14 13:58:20 +01:00
nm_clear_g_free(&priv->permissions);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: keep context-busy-watcher of NMClient alive for one more idle round The context-busy-watch has two purposes: 1) it allows the user to watch whether the NMClient still has pending GSource'es attached to the GMainContext. 2) thereby, it also keeps the inner GMainContext integrated into the caller's (in case of synchronous initialization of NMClient). Especially for 2), we must not get this wrong. Otherwise, we might un-integrate the inner GMainContext too early and it will be leaked indefinitely (because the user has no means to access or iterate it). To be extra careful, extend the lifetime of the context-busy-watcher for one more idle invocation. Theoretically, this should not be necessary, but it's not clear whether something else is still pending. The downside of that extra safety is that it is probably unnecessary in practice. And in case where it is necessary, it hides an actual issue, making it harder to notice and fix it.
2020-01-14 13:58:20 +01:00
g_clear_object(&priv->dbus_connection);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
g_clear_object(&priv->context_busy_watcher);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
nm_clear_g_free(&priv->name_owner);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: expose nml_cleanup_context_busy_watcher_on_idle() helper for reuse This can be used by NMSecretAgentOld.
2020-01-20 11:09:32 +01:00
priv->nm.capabilities_len = 0;
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_clear_g_free(&priv->nm.capabilities_arr);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_NMCLIENT_LOG_D(self, "disposed");
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
}
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface _nml_dbus_meta_iface_nm_agentmanager =
NML_DBUS_META_IFACE_INIT(NM_DBUS_INTERFACE_AGENT_MANAGER,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
NULL,
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
NML_DBUS_META_INTERFACE_PRIO_NONE, );
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
const NMLDBusMetaIface _nml_dbus_meta_iface_nm = NML_DBUS_META_IFACE_INIT_PROP(
NM_DBUS_INTERFACE,
nm_client_get_type,
NML_DBUS_META_INTERFACE_PRIO_NMCLIENT,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_DBUS_META_IFACE_DBUS_PROPERTIES(
NML_DBUS_META_PROPERTY_INIT_O_PROP(
"ActivatingConnection",
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
PROP_ACTIVATING_CONNECTION,
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NMClient,
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
_priv.nm.property_o[PROPERTY_O_IDX_NM_ACTIVATING_CONNECTION],
libnm: track permissions in NMClient as an array of well known permissions On D-Bus, the permission names are just the PolicyKit action names, like "org.freedesktop.NetworkManager.wifi.scan". But NMClient already ignores all strings that it doesn't know at compile time and only keeps track of well known permission. And neither does the API nm_client_get_permissions_result() allow to expose permissions unknown to libnm. Maybe the API of NMClient should be more generic and allow exposing any permissions announced by NetworkManager. As it is however, it's not necessary to track the permissions in a hash table. An array with fixed indices is sufficient.
2019-12-06 11:48:04 +01:00
nm_active_connection_get_type),
libnm: add NM_CLIENT_DBUS_CONNECTION property The used GDBusConnection should be configurable when creating the NMClient instance. Automatically choosing one of the g_bus_get() singletons is fine by default, but it's an unnecessary limitation.
2019-10-12 14:59:23 +02:00
NML_DBUS_META_PROPERTY_INIT_AO_PROP(
"ActiveConnections",
PROP_ACTIVE_CONNECTIONS,
libnm: log message when NMClient gets disposed This is useful as a last message to know when the instance is gone for good.
2020-01-03 13:09:37 +01:00
NMClient,
_priv.nm.property_ao[PROPERTY_AO_IDX_ACTIVE_CONNECTIONS],
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
nm_active_connection_get_type,
.notify_changed_ao = _property_ao_notify_changed_active_connections_cb),
NML_DBUS_META_PROPERTY_INIT_AO_PROP("AllDevices",
PROP_ALL_DEVICES,
NMClient,
_priv.nm.property_ao[PROPERTY_AO_IDX_ALL_DEVICES],
nm_device_get_type,
.notify_changed_ao =
_property_ao_notify_changed_all_devices_cb),
NML_DBUS_META_PROPERTY_INIT_FCN("Capabilities",
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
PROP_CAPABILITIES,
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
"au",
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_notify_update_prop_nm_capabilities, ),
NML_DBUS_META_PROPERTY_INIT_AO_PROP("Checkpoints",
PROP_CHECKPOINTS,
NMClient,
_priv.nm.property_ao[PROPERTY_AO_IDX_CHECKPOINTS],
nm_checkpoint_get_type),
NML_DBUS_META_PROPERTY_INIT_U("Connectivity",
PROP_CONNECTIVITY,
NMClient,
_priv.nm.connectivity),
NML_DBUS_META_PROPERTY_INIT_B("ConnectivityCheckAvailable",
PROP_CONNECTIVITY_CHECK_AVAILABLE,
NMClient,
_priv.nm.connectivity_check_available),
NML_DBUS_META_PROPERTY_INIT_B("ConnectivityCheckEnabled",
PROP_CONNECTIVITY_CHECK_ENABLED,
NMClient,
_priv.nm.connectivity_check_enabled),
NML_DBUS_META_PROPERTY_INIT_S("ConnectivityCheckUri",
PROP_CONNECTIVITY_CHECK_URI,
NMClient,
_priv.nm.connectivity_check_uri),
NML_DBUS_META_PROPERTY_INIT_AO_PROP("Devices",
PROP_DEVICES,
NMClient,
_priv.nm.property_ao[PROPERTY_AO_IDX_DEVICES],
nm_device_get_type,
.notify_changed_ao =
_property_ao_notify_changed_devices_cb),
NML_DBUS_META_PROPERTY_INIT_IGNORE("GlobalDnsConfiguration", "a{sv}"),
NML_DBUS_META_PROPERTY_INIT_U("Metered", PROP_METERED, NMClient, _priv.nm.metered),
NML_DBUS_META_PROPERTY_INIT_B("NetworkingEnabled",
PROP_NETWORKING_ENABLED,
NMClient,
_priv.nm.networking_enabled),
NML_DBUS_META_PROPERTY_INIT_O_PROP("PrimaryConnection",
PROP_PRIMARY_CONNECTION,
NMClient,
_priv.nm.property_o[PROPERTY_O_IDX_NM_PRIMAY_CONNECTION],
nm_active_connection_get_type),
NML_DBUS_META_PROPERTY_INIT_IGNORE("PrimaryConnectionType", "s"),
libnm: support radio flags
2022-03-21 10:20:11 +01:00
NML_DBUS_META_PROPERTY_INIT_U("RadioFlags",
PROP_RADIO_FLAGS,
NMClient,
_priv.nm.radio_flags),
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_DBUS_META_PROPERTY_INIT_B("Startup", PROP_STARTUP, NMClient, _priv.nm.startup),
NML_DBUS_META_PROPERTY_INIT_U("State", PROP_STATE, NMClient, _priv.nm.state),
NML_DBUS_META_PROPERTY_INIT_S("Version", PROP_VERSION, NMClient, _priv.nm.version),
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
NML_DBUS_META_PROPERTY_INIT_FCN("VersionInfo",
PROP_VERSION_INFO,
"au",
_notify_update_prop_nm_version_info, ),
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_DBUS_META_PROPERTY_INIT_IGNORE("WimaxEnabled", "b"),
NML_DBUS_META_PROPERTY_INIT_IGNORE("WimaxHardwareEnabled", "b"),
NML_DBUS_META_PROPERTY_INIT_B("WirelessEnabled",
PROP_WIRELESS_ENABLED,
NMClient,
_priv.nm.wireless_enabled),
NML_DBUS_META_PROPERTY_INIT_B("WirelessHardwareEnabled",
PROP_WIRELESS_HARDWARE_ENABLED,
NMClient,
_priv.nm.wireless_hardware_enabled),
NML_DBUS_META_PROPERTY_INIT_B("WwanEnabled",
PROP_WWAN_ENABLED,
NMClient,
_priv.nm.wwan_enabled),
NML_DBUS_META_PROPERTY_INIT_B("WwanHardwareEnabled",
PROP_WWAN_HARDWARE_ENABLED,
NMClient,
_priv.nm.wwan_hardware_enabled), ), );
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface _nml_dbus_meta_iface_nm_settings = NML_DBUS_META_IFACE_INIT_PROP(
NM_DBUS_INTERFACE_SETTINGS,
nm_client_get_type,
NML_DBUS_META_INTERFACE_PRIO_NMCLIENT,
NML_DBUS_META_IFACE_DBUS_PROPERTIES(
NML_DBUS_META_PROPERTY_INIT_B("CanModify",
PROP_CAN_MODIFY,
NMClient,
_priv.settings.can_modify),
NML_DBUS_META_PROPERTY_INIT_AO_PROP(
"Connections",
PROP_CONNECTIONS,
NMClient,
_priv.settings.connections,
nm_remote_connection_get_type,
.notify_changed_ao = _property_ao_notify_changed_connections_cb,
all: code format
2025-05-13 11:43:33 +02:00
.check_nmobj_visible_fcn = (gboolean (*)(GObject *)) nm_remote_connection_get_visible),
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
NML_DBUS_META_PROPERTY_INIT_S("Hostname",
PROP_HOSTNAME,
NMClient,
_priv.settings.hostname), ), );
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
const NMLDBusMetaIface _nml_dbus_meta_iface_nm_dnsmanager = NML_DBUS_META_IFACE_INIT_PROP(
NM_DBUS_INTERFACE_DNS_MANAGER,
nm_client_get_type,
NML_DBUS_META_INTERFACE_PRIO_NMCLIENT,
NML_DBUS_META_IFACE_DBUS_PROPERTIES(
NML_DBUS_META_PROPERTY_INIT_FCN("Configuration",
PROP_DNS_CONFIGURATION,
"aa{sv}",
_notify_update_prop_dns_manager_configuration),
NML_DBUS_META_PROPERTY_INIT_S("Mode", PROP_DNS_MODE, NMClient, _priv.dns_manager.mode),
NML_DBUS_META_PROPERTY_INIT_S("RcManager",
PROP_DNS_RC_MANAGER,
NMClient,
_priv.dns_manager.rc_manager), ), );
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
static void
nm_client_class_init(NMClientClass *client_class)
{
GObjectClass *object_class = G_OBJECT_CLASS(client_class);
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_dbus_path_nm = nm_ref_string_new(NM_DBUS_PATH);
_dbus_path_settings = nm_ref_string_new(NM_DBUS_PATH_SETTINGS);
_dbus_path_dns_manager = nm_ref_string_new(NM_DBUS_PATH_DNS_MANAGER);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
object_class->get_property = get_property;
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
object_class->set_property = set_property;
libnm: remember the caller's GMainContext when creating NMClient We will require this later. The NMClient shall be tied to the GMainContext at the moment when the instance gets created. This allows the user to have multiple, indendent NMClient instances (on different threads and GMainContext). Currently this is still unused, it will be later.
2019-10-12 14:17:16 +02:00
object_class->constructed = constructed;
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
object_class->dispose = dispose;
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm: add NM_CLIENT_DBUS_CONNECTION property The used GDBusConnection should be configurable when creating the NMClient instance. Automatically choosing one of the g_bus_get() singletons is fine by default, but it's an unnecessary limitation.
2019-10-12 14:59:23 +02:00
/**
* NMClient:dbus-connection:
*
* The #GDBusConnection to use.
*
* If this is not set during object construction, the D-Bus connection will
* automatically be chosen during async/sync initalization via g_bus_get().
*
* Since: 1.22
*/
obj_properties[PROP_DBUS_CONNECTION] = g_param_spec_object(
NM_CLIENT_DBUS_CONNECTION,
"",
"",
G_TYPE_DBUS_CONNECTION,
G_PARAM_READABLE | G_PARAM_WRITABLE | G_PARAM_CONSTRUCT_ONLY | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
/**
* NMClient:instance-flags:
*
* #NMClientInstanceFlags for the instance. These affect behavior of #NMClient.
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
* This is a construct property and you may only set most flags only during
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
* construction.
*
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
* The flag %NM_CLIENT_INSTANCE_FLAGS_NO_AUTO_FETCH_PERMISSIONS can be toggled any time,
* even after constructing the instance. Note that you may want to watch NMClient:permissions-state
* property to know whether permissions are ready. Note that permissions are only fetched
* when NMClient has a D-Bus name owner.
*
libnm: add "initialized-{good,bad}" flags to NMClientInstanceFlags When using async initialization with GAsyncInitable, the user usually can only know that initialization is complete by passing a callback. In simple cases, that can be cumbersome. Also expose a flag that allows to poll that information. Reuse the existing NM_CLIENT_INSTANCE_FLAGS for that. There is an ugliness here, that suddenly there are instance flags that cannot be set, but are still returned by the getter. But as this is a relatively obscure feature, it seems more lightweight to implement it this way (instead of adding a separate property and getter function).
2022-10-06 10:13:14 +02:00
* The flags %NM_CLIENT_INSTANCE_FLAGS_INITIALIZED_GOOD and %NM_CLIENT_INSTANCE_FLAGS_INITIALIZED_BAD
* cannot be set, however they will be returned by the getter after initialization completes.
*
libnm: add NMClient:instance-flags property Add a flags property to control behavior of NMClient. Possible future use cases: - currently it would always automatically fetch permissions. Often that is not used and the user could opt out of it. - currently, using sync init creates an internal GMainContext. This has an overhead and may be undesirable. We could implement another "sync" initialization that would merely iterate the callers mainloop until the initialization completes. A flag would allow to opt in. - currently, NMClient always fetches all connection settings automatically. Via a flag the user could opt out of that. Instead NMClient could provide an API so the user can request settings as they are needed.
2019-12-05 18:17:43 +01:00
* Since: 1.24
*/
obj_properties[PROP_INSTANCE_FLAGS] = g_param_spec_uint(
NM_CLIENT_INSTANCE_FLAGS,
"",
"",
0,
G_MAXUINT32,
0,
G_PARAM_READABLE | G_PARAM_WRITABLE | G_PARAM_CONSTRUCT | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add NM_CLIENT_DBUS_NAME_OWNER property It's not yet implemented. But obviously it's interesting to get the name owner to which the NMClient is currently connected. Note only that: the name-owner property really says whether NM is currently running or not.
2019-10-15 16:08:36 +02:00
/**
* NMClient:dbus-name-owner:
*
* The name owner of the NetworkManager D-Bus service.
*
* Since: 1.22
**/
obj_properties[PROP_DBUS_NAME_OWNER] =
g_param_spec_string(NM_CLIENT_DBUS_NAME_OWNER,
"",
"",
NULL,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:version:
*
* The NetworkManager version.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_VERSION] = g_param_spec_string(NM_CLIENT_VERSION,
"",
"",
NULL,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
/**
* NMClient:version-info: (type GArray(guint32))
*
* Expose version info and capabilities of NetworkManager. If non-empty,
* the first element is NM_VERSION, which encodes the version of the
* daemon as "(major << 16 | minor << 8 | micro)". The following elements
libnmc: fix bug checking VersionInfo's capabilities Remove the `+ 31u` that was making that it would search for bit 1 at array's element 1, instead of element 0. Fixed comparison >len that shoudl be >=len. Fix a few typos. Fixes: bc6098d44106 ('libnm: add internal nmc_client_has_{version_info_v,version_info_capability,capability}() helper')
2024-12-04 14:24:38 +01:00
* is a bitfields of %NMVersionInfoCapability. If a bit is set, then
core: add "VersionInfo" property on D-Bus and NMClient This exposes NM_VERSION as number (contrary to the "Version", which is a string). That is in particular useful, because the number can be compared with <> due to the encoding of the version. While at it, don't make it a single number. Expose an array of numbers, where the following numbers are a bitfield of capabilities. Note that before commit 3c67a1ec5e8e ('cli: remove version check against NM'), we used to parse the "Version" string to detect the version. As such, the information that "VersionInfo" exposes now, was already (somewhat) available, you just had to parse the string. The main benefit of "VersionInfo" is that it can expose capabilities (patched behavior) in in a lightweight bitfield. To include the numerical version there is just useful on top. Currently no additional capabilities are exposed. The idea is of course to have a place in the future, where we can expose additional capabilities. Adding a capability flag is most useful for behavior that we backport to older branches. Otherwise, we could just check the daemon version alone. But since we only add "VersionInfo" property only now, we cannot backport any capability further than this, because the "VersionInfo" property itself won't be backported. As such, this will only be useful in the future by having a place where we can add (and backport) capabilities. Note that there is some overlap with the existing "Capability" property and NMCapability enum. The difference is that adding a capability via "VersionInfo" is only one bit, and thus cheaper. Most importantly, having it cheaper means the downsides of adding a capability flag is significantly removed. In practice, we could live without capabilities for a long time, so they must be very cheap for them to be worth to add. Another difference might be, that we will want that the VersionInfo is about compile time defaults (e.g. a certain patch/behavior that is in or not), while NM_CAPABILITY_TEAM depends on whether the team plugin is loaded at runtime.
2022-12-13 12:25:04 +01:00
* the running NetworkManager has the respective capability.
*
* Since: 1.42
*/
obj_properties[PROP_VERSION_INFO] =
g_param_spec_boxed(NM_CLIENT_VERSION_INFO,
"",
"",
G_TYPE_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:state:
*
* The current daemon state.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_STATE] = g_param_spec_enum(NM_CLIENT_STATE,
"",
"",
NM_TYPE_STATE,
NM_STATE_UNKNOWN,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:startup:
*
* Whether the daemon is still starting up.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_STARTUP] = g_param_spec_boolean(NM_CLIENT_STARTUP,
"",
"",
FALSE,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
libnm: synchronize NMClient and NMRemoteSettings "is NM running" properties Rename NMClient:manager-running and NMRemoteSettings:service-running to both be :nm-running.
2014-08-05 10:05:26 -04:00
* NMClient:nm-running:
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
* Whether the daemon is running.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_NM_RUNNING] =
g_param_spec_boolean(NM_CLIENT_NM_RUNNING,
"",
"",
FALSE,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:networking-enabled:
*
* Whether networking is enabled.
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
* The property setter is a synchronous D-Bus call. This is deprecated since 1.22.
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_NETWORKING_ENABLED] =
g_param_spec_boolean(NM_CLIENT_NETWORKING_ENABLED,
"",
"",
libnm: change default value for NMClient:{networking,wireless-hardware}-enabled properties
2019-10-21 18:34:09 +02:00
FALSE,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:wireless-enabled:
*
* Whether wireless is enabled.
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
* The property setter is a synchronous D-Bus call. This is deprecated since 1.22.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_WIRELESS_ENABLED] =
g_param_spec_boolean(NM_CLIENT_WIRELESS_ENABLED,
"",
"",
FALSE,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:wireless-hardware-enabled:
*
* Whether the wireless hardware is enabled.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_WIRELESS_HARDWARE_ENABLED] =
g_param_spec_boolean(NM_CLIENT_WIRELESS_HARDWARE_ENABLED,
"",
"",
libnm: change default value for NMClient:{networking,wireless-hardware}-enabled properties
2019-10-21 18:34:09 +02:00
FALSE,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:wwan-enabled:
*
* Whether WWAN functionality is enabled.
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
* The property setter is a synchronous D-Bus call. This is deprecated since 1.22.
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_WWAN_ENABLED] =
g_param_spec_boolean(NM_CLIENT_WWAN_ENABLED,
"",
"",
FALSE,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:wwan-hardware-enabled:
*
* Whether the WWAN hardware is enabled.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_WWAN_HARDWARE_ENABLED] =
g_param_spec_boolean(NM_CLIENT_WWAN_HARDWARE_ENABLED,
"",
"",
FALSE,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:wimax-enabled:
*
* Whether WiMAX functionality is enabled.
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
* Deprecated: 1.22: WiMAX is no longer supported and this always returns FALSE. The setter has no effect.
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_WIMAX_ENABLED] =
g_param_spec_boolean(NM_CLIENT_WIMAX_ENABLED,
"",
"",
FALSE,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:wimax-hardware-enabled:
*
* Whether the WiMAX hardware is enabled.
libnm: retire nm_client_wimax_*() functions The server doesn's support WiMAX anymore. Hence there is no point in keeping this functionality. While we cannot drop the functions, let them not do anything. The code in NMManager is still there. But since we will soon drop NMManager entirely, it doesn't matter.
2019-10-29 20:05:02 +01:00
*
* Deprecated: 1.22: WiMAX is no longer supported and this always returns FALSE.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_WIMAX_HARDWARE_ENABLED] =
g_param_spec_boolean(NM_CLIENT_WIMAX_HARDWARE_ENABLED,
"",
"",
FALSE,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: support radio flags
2022-03-21 10:20:11 +01:00
/**
* NMClient:radio-flags:
*
* Flags for radio interfaces. See #NMRadioFlags.
*
* Since: 1.38
**/
obj_properties[PROP_RADIO_FLAGS] = g_param_spec_uint(NM_CLIENT_RADIO_FLAGS,
"",
"",
0,
G_MAXUINT32,
NM_RADIO_FLAG_NONE,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
gobject-introspection: made several fixes to the annotations https://bugzilla.gnome.org/show_bug.cgi?id=794658
2018-03-24 15:18:21 +00:00
* NMClient:active-connections: (type GPtrArray(NMActiveConnection))
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
* The active connections.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_ACTIVE_CONNECTIONS] =
g_param_spec_boxed(NM_CLIENT_ACTIVE_CONNECTIONS,
"",
"",
G_TYPE_PTR_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:connectivity:
*
* The network connectivity state.
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_CONNECTIVITY] =
g_param_spec_enum(NM_CLIENT_CONNECTIVITY,
"",
"",
NM_TYPE_CONNECTIVITY_STATE,
NM_CONNECTIVITY_UNKNOWN,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
/**
* NMClient::connectivity-check-available
*
* Whether a connectivity checking service has been configured.
libnm: add gtk-doc Since markers for new API and missing enum documentation
2017-08-17 23:08:44 +02:00
*
* Since: 1.10
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_CONNECTIVITY_CHECK_AVAILABLE] =
g_param_spec_boolean(NM_CLIENT_CONNECTIVITY_CHECK_AVAILABLE,
"",
"",
FALSE,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
/**
* NMClient::connectivity-check-enabled
*
* Whether a connectivity checking service has been enabled.
libnm: add gtk-doc Since markers for new API and missing enum documentation
2017-08-17 23:08:44 +02:00
*
libnm: deprecate synchronous/blocking API in libnm Note that D-Bus is fundamentally asynchronous. Doing blocking calls on top of D-Bus is odd, especially for libnm's NMClient. That is because NMClient essentially is a client-side cache of the objects from the D-Bus interface. This cache should be filled exclusively by (asynchronous) D-Bus events (PropertiesChanged). So, making a blocking D-Bus call means to wait for a response and return it, while queuing all messages that are received in the meantime. Basically there are three ways how a synchronous API on NMClient could behave: 1) the call just calls g_dbus_connection_call_sync(). This means that libnm sends a D-Bus request via GDBusConnection, and blockingly waits for the response. All D-Bus messages that get received in the meantime are queued in the GMainContext that belongs to NMClient. That means, none of these D-Bus events are processed until we iterate the GMainContext after the call returns. The effect is, that NMClient (and all cached objects in there) are unaffected by the D-Bus request. Most of the synchronous API calls in libnm are of this kind. The problem is that the strict ordering of D-Bus events gets violated. For some API this is not an immediate problem. Take for example nm_device_wifi_request_scan(). The call merely blockingly tells NetworkManager to start scanning, but since NetworkManager's D-Bus API does not directly expose any state that tells whether we are currently scanning, this out of order processing of the D-Bus request is a small issue. The problem is more obvious for nm_client_networking_set_enabled(). After calling it, NM_CLIENT_NETWORKING_ENABLED is still unaffected and unchanged, because the PropertiesChanged signal from D-Bus is not yet processed. This means, while you make such a blocking call, NMClient's state does not change. But usually you perform the synchronous call to change some state. In this form, the blocking call is not useful, because NMClient only changes the state after iterating the GMainContext, and not after the blocking call returns. 2) like 1), but after making the blocking g_dbus_connection_call_sync(), update the NMClient cache artificially. This is what nm_manager_check_connectivity() does, to "fix" bgo#784629. This also has the problem of out-of-order events, but it kinda solves the problem of not changing the state during the blocking call. But it does so by hacking the state of the cache. I think this is really wrong because the state should only be updated from the ordered stream of D-Bus messages (PropertiesChanged signal and similar). When libnm decides to modify the state, there may be already D-Bus messages queued that affect this very state. 3) instead of calling g_dbus_connection_call_sync(), use the asynchronous g_dbus_connection_call(). If we would use a sepaate GMainContext for all D-Bus related calls, we could ensure that while we block for the response, we iterate that internal main context. This might be nice, because all events are processed in order and after the blocking call returns, the NMClient state is up to date. The are problems however: current blocking API does not do this, so it's a significant change in behavior. Also, it might be unexpected to the user that during the blocking call the entire content of NMClient's cache might change and all pointers to the cache might be invalidated. Also, of course NMClient would invoke signals for all the changes that happen. Another problem is that this would be more effort to implement and it involves a small performance overhead for all D-Bus related calls (because we have to serialize all events in an internal GMainContext first and then invoke them on the caller's context). Also, if the users wants this behavior, they could implement it themself by running libnm in their own GMainContext. Note that libnm might have bugs to make that really working, but that should be fixed instead of adding such synchrnous API behavior. Read also [1], for why blocking calls are wrong. [1] https://smcv.pseudorandom.co.uk/2008/11/nonblocking/ So, all possible behaviors for synchronous API have severe behavioural issues. Mark all this API as deprecated. Also, this serves the purpose of identifying blocking D-Bus calls in libnm. Note that "deprecated" here does not really mean that the API is going to be removed. We don't break API. The user may: - continue to use this API. It's deprecated, awkward and discouraged, but if it works, by all means use it. - use asynchronous API. That's the only sensible way to use D-Bus. If libnm lacks a certain asynchronous counterpart, it should be added. - use GDBusConnection directly. There really isn't anything wrong with D-Bus or GDBusConnection. This deprecated API is just a wrapper around g_dbus_connection_call_sync(). You may call it directly without feeling dirty. --- The only other remainging API is the synchronous GInitable call for NMClient. That is an entirely separate beast and not particularly wrong (from an API point of view). Note that synchronous API in NMSecretAgentOld, NMVpnPluginOld and NMVpnServicePlugin as not deprecated here. These types are not part of the D-Bus cache and while they have similar issues, it's less severe because they have less state.
2019-09-04 13:58:43 +02:00
* The property setter is a synchronous D-Bus call. This is deprecated since 1.22.
libnm: fix placement of some "Since:" tags libnm-client-impl/nm-client.c:8398: warning: multi-line since docs found libnm-client-impl/nm-device-macvlan.c:115: warning: multi-line since docs found libnm-client-impl/nm-device-vxlan.c:540: warning: multi-line since docs found libnm-client-impl/nm-device-vxlan.c:92: warning: multi-line since docs found libnm-core-impl/nm-setting-ethtool.c:41: warning: multi-line since docs found libnm-core-impl/nm-setting-ip-config.c:2475: warning: multi-line since docs found libnm-core-impl/nm-setting-ip-config.c:2504: warning: multi-line since docs found
2022-05-05 12:04:36 +02:00
*
* Since: 1.10
client: expose connectivity-check-{available,enabled} properties https://bugzilla.gnome.org/show_bug.cgi?id=785117
2017-08-09 15:21:28 +08:00
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_CONNECTIVITY_CHECK_ENABLED] =
g_param_spec_boolean(NM_CLIENT_CONNECTIVITY_CHECK_ENABLED,
"",
"",
FALSE,
G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
/**
* NMClient:connectivity-check-uri:
*
* The used URI for connectivity checking.
*
* Since: 1.22
**/
obj_properties[PROP_CONNECTIVITY_CHECK_URI] =
g_param_spec_string(NM_CLIENT_CONNECTIVITY_CHECK_URI,
"",
"",
NULL,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:primary-connection:
*
* The #NMActiveConnection of the device with the default route;
* see nm_client_get_primary_connection() for more details.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_PRIMARY_CONNECTION] =
g_param_spec_object(NM_CLIENT_PRIMARY_CONNECTION,
"",
"",
NM_TYPE_ACTIVE_CONNECTION,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient:activating-connection:
*
* The #NMActiveConnection of the activating connection that is
* likely to become the new #NMClient:primary-connection.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_ACTIVATING_CONNECTION] =
g_param_spec_object(NM_CLIENT_ACTIVATING_CONNECTION,
"",
"",
NM_TYPE_ACTIVE_CONNECTION,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
gobject-introspection: made several fixes to the annotations https://bugzilla.gnome.org/show_bug.cgi?id=794658
2018-03-24 15:18:21 +00:00
* NMClient:devices: (type GPtrArray(NMDevice))
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
*
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
* List of real network devices. Does not include placeholder devices.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_DEVICES] = g_param_spec_boxed(NM_CLIENT_DEVICES,
"",
"",
G_TYPE_PTR_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
/**
gobject-introspection: made several fixes to the annotations https://bugzilla.gnome.org/show_bug.cgi?id=794658
2018-03-24 15:18:21 +00:00
* NMClient:all-devices: (type GPtrArray(NMDevice))
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
*
* List of both real devices and device placeholders.
* Since: 1.2
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_ALL_DEVICES] =
g_param_spec_boxed(NM_CLIENT_ALL_DEVICES,
"",
"",
G_TYPE_PTR_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
gobject-introspection: made several fixes to the annotations https://bugzilla.gnome.org/show_bug.cgi?id=794658
2018-03-24 15:18:21 +00:00
* NMClient:connections: (type GPtrArray(NMRemoteConnection))
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
*
* The list of configured connections that are available to the user. (Note
* that this differs from the underlying D-Bus property, which may also
* contain the object paths of connections that the user does not have
* permission to read the details of.)
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_CONNECTIONS] =
g_param_spec_boxed(NM_CLIENT_CONNECTIONS,
"",
"",
G_TYPE_PTR_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
* NMClient:hostname:
*
* The machine hostname stored in persistent configuration. This can be
* modified by calling nm_client_save_hostname().
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_HOSTNAME] = g_param_spec_string(NM_CLIENT_HOSTNAME,
"",
"",
NULL,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
* NMClient:can-modify:
*
* If %TRUE, adding and modifying connections is supported.
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_CAN_MODIFY] =
g_param_spec_boolean(NM_CLIENT_CAN_MODIFY,
"",
"",
FALSE,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
nm-manager: add 'metered' property This introduces a global metered property which makes easier for clients to obtain the metered status of the current primary connection.
2015-06-03 09:15:24 +02:00
/**
* NMClient:metered:
*
* Whether the connectivity is metered.
*
* Since: 1.2
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_METERED] = g_param_spec_uint(NM_CLIENT_METERED,
"",
"",
0,
G_MAXUINT32,
NM_METERED_UNKNOWN,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
/**
* NMClient:dns-mode:
*
* The current DNS processing mode.
*
* Since: 1.6
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_DNS_MODE] = g_param_spec_string(NM_CLIENT_DNS_MODE,
"",
"",
libnm: change default value for NMClient:dns-{mode,rc-manager} properties
2019-10-21 18:34:09 +02:00
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
/**
* NMClient:dns-rc-manager:
*
* The current resolv.conf management mode.
*
* Since: 1.6
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_DNS_RC_MANAGER] =
g_param_spec_string(NM_CLIENT_DNS_RC_MANAGER,
"",
"",
libnm: change default value for NMClient:dns-{mode,rc-manager} properties
2019-10-21 18:34:09 +02:00
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
/**
libnm: fix dns-configuration property type Fixes: a8d600525643 ('libnm: implement support for DNS manager properties') https://bugzilla.redhat.com/show_bug.cgi?id=1689055 (cherry picked from commit d867837d05a808ff90756a41a96f778846522465)
2019-03-15 09:41:51 +01:00
* NMClient:dns-configuration: (type GPtrArray(NMDnsEntry))
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
*
libnm: fix dns-configuration property type Fixes: a8d600525643 ('libnm: implement support for DNS manager properties') https://bugzilla.redhat.com/show_bug.cgi?id=1689055 (cherry picked from commit d867837d05a808ff90756a41a96f778846522465)
2019-03-15 09:41:51 +01:00
* The current DNS configuration, represented as an array
* of #NMDnsEntry objects.
libnm: implement support for DNS manager properties
2016-10-25 11:11:12 +02:00
*
* Since: 1.6
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_DNS_CONFIGURATION] =
g_param_spec_boxed(NM_CLIENT_DNS_CONFIGURATION,
"",
"",
G_TYPE_PTR_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
/**
gobject-introspection: made several fixes to the annotations https://bugzilla.gnome.org/show_bug.cgi?id=794658
2018-03-24 15:18:21 +00:00
* NMClient:checkpoints: (type GPtrArray(NMCheckpoint))
libnm: add checkpoint support
2017-10-21 16:05:19 +02:00
*
* The list of active checkpoints.
*
* Since: 1.12
*/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
obj_properties[PROP_CHECKPOINTS] =
libnm: add NM_CLIENT_CHECKPOINTS define
2019-10-22 14:17:11 +02:00
g_param_spec_boxed(NM_CLIENT_CHECKPOINTS,
"",
"",
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_TYPE_PTR_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
/**
* NMClient:capabilities: (type GArray(guint32))
*
* The list of capabilities numbers as guint32 or %NULL if
all: fix minor typos https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/565
2020-07-04 11:37:01 +03:00
* there are no capabilities. The numeric value correspond
libnm: add nm_client_get_capabilities() to expose server Capabilities I hesitated to add this to libnm, because it's hardly used. However, we already fetch the property during GetManagedObjects(), we we should make it accessible, instead of requiring the user to make another D-Bus call.
2019-12-20 18:10:55 +01:00
* to %NMCapability enum.
*
* Since: 1.24
*/
obj_properties[PROP_CAPABILITIES] =
g_param_spec_boxed(NM_CLIENT_CAPABILITIES,
"",
"",
G_TYPE_ARRAY,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: allow to enable/disable fetching of permissions in NMClient Currently, NMClient by default always fetches the permissions ("GetPermissions()") and refreshes them on "CheckPermissions" signal. Fetching permissions is relatively expensive, while they are not used most of the time. Allow the user to opt out of this. For that, have a NMClientInstanceFlags to enable/disable automatic fetching. Also add a "permissions-state" property that allows the user to understand whether the cached permissions are up to date or not. This is a bit an awkward API for handling this. E.g. you cannot explicitly request permissions, you can just enable/disable fetching permissions. And then you can watch the permission-state to know whether you are ready. It's done this way because it fits the previous model and extends the API with a (relative) small amount of new functions and properties.
2019-12-05 15:53:51 +01:00
/**
* NMClient:permissions-state:
*
* The state of the cached permissions. The value %NM_TERNARY_DEFAULT
* means that no permissions are yet received (or not yet requested).
* %NM_TERNARY_TRUE means that permissions are received, cached and up
* to date. %NM_TERNARY_FALSE means that permissions were received and are
* cached, but in the meantime a "CheckPermissions" signal was received
* that invalidated the cached permissions.
* Note that NMClient will always emit a notify::permissions-state signal
* when a "CheckPermissions" signal got received or after new permissions
* got received (that is regardless whether the value of the permission state
* actually changed). With this you can watch the permissions-state property
* to know whether the permissions are ready. Note that while NMClient has
* no D-Bus name owner, no permissions are fetched (and this property won't
* change).
*
* Since: 1.24
*/
obj_properties[PROP_PERMISSIONS_STATE] =
g_param_spec_enum(NM_CLIENT_PERMISSIONS_STATE,
"",
"",
NM_TYPE_TERNARY,
NM_TERNARY_DEFAULT,
G_PARAM_READABLE | G_PARAM_STATIC_STRINGS);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: refactor caching of D-Bus objects in NMClient No longer use GDBusObjectMangaerClient and gdbus-codegen generated classes for the NMClient cache. Instead, use GDBusConnection directly and a custom implementation (NMLDBusObject) for caching D-Bus' ObjectManager data. CHANGES ------- - This is a complete rework. I think the previous implementation was difficult to understand. There were unfixed bugs and nobody understood the code well enough to fix them. Maybe somebody out there understood the code, but I certainly did not. At least nobody provided patches to fix those issues. I do believe that this implementation is more straightforward and easier to understand. It removes a lot of layers of code. Whether this claim of simplicity is true, each reader must decide for himself/herself. Note that it is still fairly complex. - There was a lingering performance issue with large number of D-Bus objects. The patch tries hard that the implementation scales well. Of course, when we cache N objects that have N-to-M references to other, we still are fundamentally O(N*M) for runtime and memory consumption (with M being the number of references between objects). But each part should behave efficiently and well. - Play well with GMainContext. libnm code (NMClient) is generally not thread safe. However, it should work to use multiple instances in parallel, as long as each access to a NMClient is through the caller's GMainContext. This follows glib's style and effectively allows to use NMClient in a multi threaded scenario. This implies to stick to a main context upon construction and ensure that callbacks are only invoked when iterating that context. Also, NMClient itself shall never iterate the caller's context. This also means, libnm must never use g_idle_add() or g_timeout_add(), as those enqueue sources in the g_main_context_default() context. - Get ordering of messages right. All events are consistently enqueued in a GMainContext and processed strictly in order. For example, previously "nm-object.c" tried to combine signals and emit them on an idle handler. That is wrong, signals must be emitted in the right order and when they happen. Note that when using GInitable's synchronous initialization to initialize the NMClient instance, NMClient internally still operates fully asynchronously. In that case NMClient has an internal main context. - NMClient takes over most of the functionality. When using D-Bus' ObjectManager interface, one needs to handle basically the entire state of the D-Bus interface. That cannot be separated well into distinct parts, and even if you try, you just end up having closely related code in different source files. Spreading related code does not make it easier to understand, on the contrary. That means, NMClient is inherently complex as it contains most of the logic. I think that is not avoidable, but it's not as bad as it sounds. - NMClient processes D-Bus messages and state changes in separate steps. First NMClient unpacks the message (e.g. _dbus_handle_properties_changed()) and keeps track of the changed data. Then we update the GObject instances (_dbus_handle_obj_changed_dbus()) without emitting any signals yet. Finally, we emit all signals and notifications that were collected (_dbus_handle_changes_commit()). Note that for example during the initial GetManagedObjects() reply, NMClient receive a large amount of state at once. But we first apply all the changes to our GObject instances before emitting any signals. The result is that signals are always emitted in a moment when the cache is consistent. The unavoidable downside is that when you receive a property changed signal, possibly many other properties changed already and more signals are about to be emitted. - NMDeviceWifi no longer modifies the content of the cache from client side during poke_wireless_devices_with_rf_status(). The content of the cache should be determined by D-Bus alone and follow what NetworkManager service exposes. Local modifications should be avoided. - This aims to bring no API/ABI change, though it does of course bring various subtle changes in behavior. Those should be all for the better, but the goal is not to break any existing clients. This does change internal (albeit externally visible) API, like dropping NM_OBJECT_DBUS_OBJECT_MANAGER property and NMObject no longer implementing GInitableIface and GAsyncInitableIface. - Some uses of gdbus-codegen classes remain in NMVpnPluginOld, NMVpnServicePlugin and NMSecretAgentOld. These are independent of NMClient/NMObject and should be reworked separately. - While we no longer use generated classes from gdbus-codegen, we don't need more glue code than before. Also before we constructed NMPropertiesInfo and a had large amount of code to propagate properties from NMDBus* to NMObject. That got completely reworked, but did not fundamentally change. You still need about the same effort to create the NMLDBusMetaIface. Not using generated bindings did not make anything worse (which tells about the usefulness of generated code, at least in the way it was used). - NMLDBusMetaIface and other meta data is static and immutable. This avoids copying them around. Also, macros like NML_DBUS_META_PROPERTY_INIT_U() have compile time checks to ensure the property types matches. It's pretty hard to misuse them because it won't compile. - The meta data now explicitly encodes the expected D-Bus types and makes sure never to accept wrong data. That would only matter when the server (accidentally or intentionally) exposes unexpected types on D-Bus. I don't think that was previously ensured in all cases. For example, demarshal_generic() only cared about the GObject property type, it didn't know the expected D-Bus type. - Previously GDBusObjectManager would sometimes emit warnings (g_log()). Those probably indicated real bugs. In any case, it prevented us from running CI with G_DEBUG=fatal-warnings, because there would be just too many unrelated crashes. Now we log debug messages that can be enabled with "LIBNM_CLIENT_DEBUG=trace". Some of these messages can also be turned into g_warning()/g_critical() by setting LIBNM_CLIENT_DEBUG=warning,error. Together with G_DEBUG=fatal-warnings, this turns them into assertions. Note that such "assertion failures" might also happen because of a server bug (or change). Thus these are not common assertions that indicate a bug in libnm and are thus not armed unless explicitly requested. In our CI we should now always run with LIBNM_CLIENT_DEBUG=warning,error and G_DEBUG=fatal-warnings and to catch bugs. Note that currently NetworkManager has bugs in this regard, so enabling this will result in assertion failures. That should be fixed first. - Note that this changes the order in which we emit "notify:devices" and "device-added" signals. I think it makes the most sense to emit first "device-removed", then "notify:devices", and finally "device-added" signals. This changes behavior for commit 52ae28f6e5bf ('libnm: queue added/removed signals and suppress uninitialized notifications'), but I don't think that users should actually rely on the order. Still, the new order makes the most sense to me. - In NetworkManager, profiles can be invisible to the user by setting "connection.permissions". Such profiles would be hidden by NMClient's nm_client_get_connections() and their "connection-added"/"connection-removed" signals. Note that NMActiveConnection's nm_active_connection_get_connection() and NMDevice's nm_device_get_available_connections() still exposes such hidden NMRemoteConnection instances. This behavior was preserved. NUMBERS ------- I compared 3 versions of libnm. [1] 962297f9085d, current tip of nm-1-20 branch [2] 4fad8c7c642e, current master, immediate parent of this patch [3] this patch All tests were done on Fedora 31, x86_64, gcc 9.2.1-1.fc31. The libraries were build with $ ./contrib/fedora/rpm/build_clean.sh -g -w test -W debug Note that RPM build already stripped the library. --- N1) File size of libnm.so.0.1.0 in bytes. There currently seems to be a issue on Fedora 31 generating wrong ELF notes. Usually, libnm is smaller but in these tests it had large (and bogus) ELF notes. Anyway, the point is to show the relative sizes, so it doesn't matter). [1] 4075552 (102.7%) [2] 3969624 (100.0%) [3] 3705208 ( 93.3%) --- N2) `size /usr/lib64/libnm.so.0.1.0`: text data bss dec hex filename [1] 1314569 (102.0%) 69980 ( 94.8%) 10632 ( 80.4%) 1395181 (101.4%) 1549ed /usr/lib64/libnm.so.0.1.0 [2] 1288410 (100.0%) 73796 (100.0%) 13224 (100.0%) 1375430 (100.0%) 14fcc6 /usr/lib64/libnm.so.0.1.0 [3] 1229066 ( 95.4%) 65248 ( 88.4%) 13400 (101.3%) 1307714 ( 95.1%) 13f442 /usr/lib64/libnm.so.0.1.0 --- N3) Performance test with test-client.py. With checkout of [2], run ``` prepare_checkout() { rm -rf /tmp/nm-test && \ git checkout -B test 4fad8c7c642e && \ git clean -fdx && \ ./autogen.sh --prefix=/tmp/nm-test && \ make -j 5 install && \ make -j 5 check-local-clients-tests-test-client } prepare_test() { NM_TEST_REGENERATE=1 NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v } do_test() { for i in {1..10}; do NM_TEST_CLIENT_BUILDDIR="/data/src/NetworkManager" NM_TEST_CLIENT_NMCLI_PATH=/usr/bin/nmcli python3 ./clients/tests/test-client.py -v || return -1 done echo "done!" } prepare_checkout prepare_test time do_test ``` [1] real 2m14.497s (101.3%) user 5m26.651s (100.3%) sys 1m40.453s (101.4%) [2] real 2m12.800s (100.0%) user 5m25.619s (100.0%) sys 1m39.065s (100.0%) [3] real 1m54.915s ( 86.5%) user 4m18.585s ( 79.4%) sys 1m32.066s ( 92.9%) --- N4) Performance. Run NetworkManager from build [2] and setup a large number of profiles (551 profiles and 515 devices, mostly unrealized). This setup is already at the edge of what NetworkManager currently can handle. Of course, that is a different issue. Here we just check how long plain `nmcli` takes on the system. ``` do_cleanup() { for UUID in $(nmcli -g NAME,UUID connection show | sed -n 's/^xx-c-.*:\([^:]\+\)$/\1/p'); do nmcli connection delete uuid "$UUID" done for DEVICE in $(nmcli -g DEVICE device status | grep '^xx-i-'); do nmcli device delete "$DEVICE" done } do_setup() { do_cleanup for i in {1..30}; do nmcli connection add type bond autoconnect no con-name xx-c-bond-$i ifname xx-i-bond-$i ipv4.method disabled ipv6.method ignore for j in $(seq $i 30); do nmcli connection add type vlan autoconnect no con-name xx-c-vlan-$i-$j vlan.id $j ifname xx-i-vlan-$i-$j vlan.parent xx-i-bond-$i ipv4.method disabled ipv6.method ignore done done systemctl restart NetworkManager.service sleep 5 } do_test() { perf stat -r 50 -B nmcli 1>/dev/null } do_test ``` [1] Performance counter stats for 'nmcli' (50 runs): 456.33 msec task-clock:u # 1.093 CPUs utilized ( +- 0.44% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,900 page-faults:u # 0.013 M/sec ( +- 0.02% ) 1,408,675,453 cycles:u # 3.087 GHz ( +- 0.48% ) 1,594,741,060 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 368,744,018 branches:u # 808.061 M/sec ( +- 0.02% ) 4,566,058 branch-misses:u # 1.24% of all branches ( +- 0.76% ) 0.41761 +- 0.00282 seconds time elapsed ( +- 0.68% ) [2] Performance counter stats for 'nmcli' (50 runs): 477.99 msec task-clock:u # 1.088 CPUs utilized ( +- 0.36% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 5,948 page-faults:u # 0.012 M/sec ( +- 0.03% ) 1,471,133,482 cycles:u # 3.078 GHz ( +- 0.36% ) 1,655,275,369 instructions:u # 1.13 insn per cycle ( +- 0.02% ) 382,595,152 branches:u # 800.433 M/sec ( +- 0.02% ) 4,746,070 branch-misses:u # 1.24% of all branches ( +- 0.49% ) 0.43923 +- 0.00242 seconds time elapsed ( +- 0.55% ) [3] Performance counter stats for 'nmcli' (50 runs): 352.36 msec task-clock:u # 1.027 CPUs utilized ( +- 0.32% ) 0 context-switches:u # 0.000 K/sec 0 cpu-migrations:u # 0.000 K/sec 4,790 page-faults:u # 0.014 M/sec ( +- 0.26% ) 1,092,341,186 cycles:u # 3.100 GHz ( +- 0.26% ) 1,209,045,283 instructions:u # 1.11 insn per cycle ( +- 0.02% ) 281,708,462 branches:u # 799.499 M/sec ( +- 0.01% ) 3,101,031 branch-misses:u # 1.10% of all branches ( +- 0.61% ) 0.34296 +- 0.00120 seconds time elapsed ( +- 0.35% ) --- N5) same setup as N4), but run `PAGER= /bin/time -v nmcli`: [1] Command being timed: "nmcli" User time (seconds): 0.42 System time (seconds): 0.04 Percent of CPU this job got: 107% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.43 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34456 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6128 Voluntary context switches: 1298 Involuntary context switches: 1106 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [2] Command being timed: "nmcli" User time (seconds): 0.44 System time (seconds): 0.04 Percent of CPU this job got: 108% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.44 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 34452 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 6169 Voluntary context switches: 1849 Involuntary context switches: 142 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 [3] Command being timed: "nmcli" User time (seconds): 0.32 System time (seconds): 0.02 Percent of CPU this job got: 102% Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.34 Average shared text size (kbytes): 0 Average unshared data size (kbytes): 0 Average stack size (kbytes): 0 Average total size (kbytes): 0 Maximum resident set size (kbytes): 29196 Average resident set size (kbytes): 0 Major (requiring I/O) page faults: 0 Minor (reclaiming a frame) page faults: 5059 Voluntary context switches: 919 Involuntary context switches: 685 Swaps: 0 File system inputs: 0 File system outputs: 0 Socket messages sent: 0 Socket messages received: 0 Signals delivered: 0 Page size (bytes): 4096 Exit status: 0 --- N6) same setup as N4), but run `nmcli monitor` and look at `ps aux` for the RSS size. USER PID %CPU %MEM VSZ RSS TTY STAT START TIME COMMAND [1] me 1492900 21.0 0.2 461348 33248 pts/10 Sl+ 15:02 0:00 nmcli monitor [2] me 1490721 5.0 0.2 461496 33548 pts/10 Sl+ 15:00 0:00 nmcli monitor [3] me 1495801 16.5 0.1 459476 28692 pts/10 Sl+ 15:04 0:00 nmcli monitor
2019-10-30 11:42:58 +01:00
_nml_dbus_meta_class_init_with_properties(object_class,
&_nml_dbus_meta_iface_nm,
&_nml_dbus_meta_iface_nm_settings,
&_nml_dbus_meta_iface_nm_dnsmanager);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient::device-added:
* @client: the client that received the signal
* @device: (type NMDevice): the new device
*
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
* Notifies that a #NMDevice is added. This signal is not emitted for
* placeholder devices.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[DEVICE_ADDED] = g_signal_new(NM_CLIENT_DEVICE_ADDED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
0,
NULL,
NULL,
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_TYPE_NONE,
1,
G_TYPE_OBJECT);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient::device-removed:
* @client: the client that received the signal
* @device: (type NMDevice): the removed device
*
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
* Notifies that a #NMDevice is removed. This signal is not emitted for
* placeholder devices.
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[DEVICE_REMOVED] = g_signal_new(NM_CLIENT_DEVICE_REMOVED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
0,
NULL,
NULL,
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_TYPE_NONE,
1,
G_TYPE_OBJECT);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
/**
* NMClient::any-device-added:
* @client: the client that received the signal
* @device: (type NMDevice): the new device
*
* Notifies that a #NMDevice is added. This signal is emitted for both
* regular devices and placeholder devices.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[ANY_DEVICE_ADDED] = g_signal_new(NM_CLIENT_ANY_DEVICE_ADDED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
0,
NULL,
NULL,
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_TYPE_NONE,
1,
G_TYPE_OBJECT);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm/libnm-glib: add NMClient.get_all_devices() method and AllDevices property Mirror new NetworkManager API to return both real devices and device placeholders.
2014-10-10 14:28:49 -05:00
/**
* NMClient::any-device-removed:
* @client: the client that received the signal
* @device: (type NMDevice): the removed device
*
* Notifies that a #NMDevice is removed. This signal is emitted for both
* regular devices and placeholder devices.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[ANY_DEVICE_REMOVED] = g_signal_new(NM_CLIENT_ANY_DEVICE_REMOVED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
0,
NULL,
NULL,
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_TYPE_NONE,
1,
G_TYPE_OBJECT);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
/**
* NMClient::permission-changed:
* @client: the client that received the signal
* @permission: a permission from #NMClientPermission
* @result: the permission's result, one of #NMClientPermissionResult
*
* Notifies that a permission has changed
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[PERMISSION_CHANGED] = g_signal_new(NM_CLIENT_PERMISSION_CHANGED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
0,
NULL,
NULL,
NULL,
G_TYPE_NONE,
2,
G_TYPE_UINT,
G_TYPE_UINT);
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
* NMClient::connection-added:
libnm-client: fix some comments Correct the mismatched arguments and descriptions, likely a copy & paste error.
2022-05-03 17:31:40 +02:00
* @client: the client that received the signal
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* @connection: the new connection
*
* Notifies that a #NMConnection has been added.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[CONNECTION_ADDED] = g_signal_new(NM_CLIENT_CONNECTION_ADDED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
0,
NULL,
NULL,
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_TYPE_NONE,
1,
NM_TYPE_REMOTE_CONNECTION);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
/**
* NMClient::connection-removed:
libnm-client: fix some comments Correct the mismatched arguments and descriptions, likely a copy & paste error.
2022-05-03 17:31:40 +02:00
* @client: the client that received the signal
libnm: merge NMRemoteSettings into NMClient Make NMRemoteSettings internal and have NMClient wrap all of its APIs, just like it does with NMManager.
2014-09-29 10:58:16 -04:00
* @connection: the removed connection
*
* Notifies that a #NMConnection has been removed.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[CONNECTION_REMOVED] = g_signal_new(NM_CLIENT_CONNECTION_REMOVED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
libnm: hide NMClient struct from public headers and use direct private field Having the NMClient/NMClientClass structs in the public header allows the user to subclass these types. Subclassing this type was never intended, nor is it supported, nor does it seem useful. Subclassing only makes sense if the type has suitable hooks to extend the type in a meaningful way. NMClient hasn't, and everybody trying to derive from this class would better delegate the actions. Also, having these structs in the public header prevents us from embedding the private data in the object structure itself. It has thus an runtime overhead and is less convenient for debugging (it's hard to find the private data pointer in gdb). Most importantly, there is no easy way to find the offset of the private data fields, short of calling NM_CLIENT_GET_PRIVATE() -- which currently is a macro. Later we want to generically lookup the offset of the private data, we would need NM_CLIENT_GET_PRIVATE() as a function. Instead, by having an internally, statically known offset, we can use that offset instead. Also drop all signal hooks. They are also not useful. This is an ABI and API change, but of something that we never wanted to be part of the ABI/API, and which hopefull nobody is using.
2019-10-18 07:44:31 +02:00
0,
NULL,
NULL,
NULL,
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
G_TYPE_NONE,
1,
NM_TYPE_REMOTE_CONNECTION);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm/client: emit signals when active connection disappears It allows us to reliably track failures to activate a connection.
2016-07-07 11:35:19 +02:00
/**
* NMClient::active-connection-added:
libnm-client: fix some comments Correct the mismatched arguments and descriptions, likely a copy & paste error.
2022-05-03 17:31:40 +02:00
* @client: the client that received the signal
libnm/client: emit signals when active connection disappears It allows us to reliably track failures to activate a connection.
2016-07-07 11:35:19 +02:00
* @active_connection: the new active connection
*
* Notifies that a #NMActiveConnection has been added.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[ACTIVE_CONNECTION_ADDED] = g_signal_new(NM_CLIENT_ACTIVE_CONNECTION_ADDED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
0,
NULL,
NULL,
NULL,
G_TYPE_NONE,
1,
NM_TYPE_ACTIVE_CONNECTION);
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm/client: emit signals when active connection disappears It allows us to reliably track failures to activate a connection.
2016-07-07 11:35:19 +02:00
/**
* NMClient::active-connection-removed:
libnm-client: fix some comments Correct the mismatched arguments and descriptions, likely a copy & paste error.
2022-05-03 17:31:40 +02:00
* @client: the client that received the signal
libnm/client: emit signals when active connection disappears It allows us to reliably track failures to activate a connection.
2016-07-07 11:35:19 +02:00
* @active_connection: the removed active connection
*
* Notifies that a #NMActiveConnection has been removed.
**/
libnm: use obj_properties array in libnm and cleanup This is not merely cosmetic. I will need the obj_properties array to lookup GParamSpec by their PROP_* enum value. The alternative would be lookup by name, which is more expensive.
2019-10-10 17:57:30 +02:00
signals[ACTIVE_CONNECTION_REMOVED] = g_signal_new(NM_CLIENT_ACTIVE_CONNECTION_REMOVED,
G_OBJECT_CLASS_TYPE(object_class),
G_SIGNAL_RUN_FIRST,
0,
NULL,
NULL,
NULL,
G_TYPE_NONE,
1,
NM_TYPE_ACTIVE_CONNECTION);
libnm: add libnm/libnm-core (part 1) This commit begins creating the new "libnm", which will replace libnm-util and libnm-glib. The main reason for the libnm-util/libnm-glib split is that the daemon needs to link to libnm-util (to get NMSettings, NMConnection, etc), but can't link to libnm-glib (because it uses many of the same type names as the NetworkManager daemon. eg, NMDevice). So the daemon links to only libnm-util, but basically all clients link to both. With libnm, there will be only a single client-visible library, and NetworkManager will internally link against a private "libnm-core" containing the parts that used to be in libnm-util. (The "libnm-core" parts still need to be in their own directory so that the daemon can see those header files without also seeing the ones in libnm/ that conflict with its own headers.) [This commit just copies the source code from libnm-util/ to libnm-core/, and libnm-glib/ to libnm/: mkdir -p libnm-core/tests/ mkdir -p libnm/tests/ cp libnm-util/*.[ch] libnm-util/nm-version.h.in libnm-core/ rm -f libnm-core/nm-version.h libnm-core/nm-setting-template.[ch] libnm-core/nm-utils-enum-types.[ch] cp libnm-util/tests/*.[ch] libnm-core/tests/ cp libnm-glib/*.[ch] libnm/ rm -f libnm/libnm_glib.[ch] libnm/libnm-glib-test.c libnm/nm-glib-enum-types.[ch] cp libnm-glib/tests/*.[ch] libnm/tests/ ]
2014-07-24 08:53:33 -04:00
}
static void
nm_client_initable_iface_init(GInitableIface *iface)
{
iface->init = init_sync;
}
static void
nm_client_async_initable_iface_init(GAsyncInitableIface *iface)
{
iface->init_async = init_async;
iface->init_finish = init_finish;
}
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
libnm: add nm_client_wait_shutdown() function for cleaning up NMClient Add a fire-and-forget function to wait for shutdown to be complete. It's not entirely trivial to ensure all resources of NMClient are cleaned up. That matters only if NMClient uses a temporary GMainContext that the user wants to release while the application continues. For example, to do some short-lived operations an a worker thread. It's not trivial also because glib provides no convenient API to integrate a GMainContext in another GMainContext. We have that code as nm_utils_g_main_context_create_integrate_source(), so add a helper function to allow the user to do this. The function allows to omit the callback, in which case the caller wouldn't know when shutdown is complete. That would still be useful however, when integrating the client's context into the caller's context, so that the client's context gets automatically iterated until completion. The following test script will run out of file descriptors, when wait_shutdown() is not used: #!/bin/python import gi gi.require_version("NM", "1.0") from gi.repository import NM, GLib for i in range(1200): print(f">>>{i}") ctx = GLib.MainContext() ctx.push_thread_default() nmc = NM.Client.new() ctx.pop_thread_default() def cb(unused, result, i): try: NM.Client.wait_shutdown_finish(result) except Exception: # cannot happen assert False else: print(f">>>>> {i} complete") nmc.wait_shutdown(True, None, cb, i) while GLib.MainContext.default().iteration(False): pass
2022-10-05 12:22:51 +02:00
/*****************************************************************************/
typedef struct {
GCancellable *cancellable;
GSource *integration_source;
GTask *task;
GSource *idle_source;
/* A weakref to nm_client_get_context_busy_watcher() */
GWeakRef weak_ref;
gulong cancellable_id;
guint64 log_ptr;
int result;
} WaitShutdownData;
G_LOCK_DEFINE_STATIC(wait_shutdown_mutex);
static NM_CACHED_QUARK_FCN("nm.client.wait-shutdown", _wait_shutdown_get_quark);
static void
_wait_shutdown_data_free(gpointer user_data)
{
WaitShutdownData *data = user_data;
nm_g_slice_free(data);
}
static gboolean
_wait_shutdown_idle_cb(gpointer user_data)
{
WaitShutdownData *data = user_data;
gs_unref_object GTask *task = NULL;
int result;
nm_clear_g_source_inst(&data->idle_source);
task = g_steal_pointer(&data->task);
result = g_atomic_int_get(&data->result);
nm_assert(NM_IN_SET(result, 0, 1));
NML_DBUS_LOG(_NML_NMCLIENT_LOG_LEVEL_COERCE(NML_DBUS_LOG_LEVEL_TRACE),
"nmclient[" NM_HASH_OBFUSCATE_PTR_FMT
"]: wait-shutdown (" NM_HASH_OBFUSCATE_PTR_FMT ")"
"%s",
data->log_ptr,
NM_HASH_OBFUSCATE_PTR(data),
!result ? " cancelled" : " completed");
if (!result) {
GError *error = NULL;
nm_utils_error_set_cancelled(&error, FALSE, NULL);
g_task_return_error(task, error);
} else
g_task_return_boolean(task, TRUE);
return G_SOURCE_CONTINUE;
}
static void
_wait_shutdown_data_clear(WaitShutdownData *data, gboolean result)
{
gs_unref_object GObject *context_busy_watcher = NULL;
if (!g_atomic_int_compare_and_exchange(&data->result, -1, !!result)) {
/* There was a race and the result is already provided. Nothing to
* do, except, if "result" indicates cancellation (FALSE), set data->result
* to FALSE to. Aside that, the completion is already in progress. */
if (!result)
g_atomic_int_compare_and_exchange(&data->result, TRUE, FALSE);
return;
}
nm_clear_g_signal_handler(data->cancellable, &data->cancellable_id);
nm_clear_g_source_inst(&data->integration_source);
g_clear_object(&data->cancellable);
if (!result) {
/* This was a cancellation. We likely still have the qdata tracked.
* We need to remove it. */
context_busy_watcher = g_weak_ref_get(&data->weak_ref);
if (context_busy_watcher) {
GPtrArray *qdata_arr;
G_LOCK(wait_shutdown_mutex);
qdata_arr = g_object_get_qdata(context_busy_watcher, _wait_shutdown_get_quark());
if (qdata_arr && g_ptr_array_remove_fast(qdata_arr, data)) {
/* data->task had an additional reference, we return it now. */
g_object_unref(data->task);
}
G_UNLOCK(wait_shutdown_mutex);
}
}
g_weak_ref_clear(&data->weak_ref);
/* We don't complete right away, instead always schedule an idle action
* on the caller's context. */
data->idle_source = nm_g_source_attach(
nm_g_idle_source_new(G_PRIORITY_DEFAULT_IDLE, _wait_shutdown_idle_cb, data, NULL),
g_task_get_context(data->task));
}
static void
_wait_shutdown_qdata_cb(gpointer user_data)
{
gs_unref_ptrarray GPtrArray *qdata_arr = user_data;
while (qdata_arr->len > 0) {
WaitShutdownData *data;
data = g_ptr_array_remove_index_fast(qdata_arr, qdata_arr->len - 1);
_wait_shutdown_data_clear(data, TRUE);
/* data->task had an additional reference, we return it now. */
g_object_unref(data->task);
}
}
static void
_wait_shutdown_cancelled_cb(GCancellable *cancellable, gpointer user_data)
{
_wait_shutdown_data_clear(g_task_get_task_data(user_data), FALSE);
}
/**
* nm_client_wait_shutdown:
* @client: the #NMClient to shutdown.
* @integrate_maincontext: whether to hook the client's maincontext
* in the current thread default. Otherwise, you must ensure
* that the client's maincontext gets iterated so that it can complete.
* By integrating the maincontext in the current thread default, you
* may instead only iterate the latter.
doc: replace all (allow-none) annotations by (optional) and/or (nullable) The (allow-none) annotation is deprecated since a long time now, it is better to use (nullable) and/or (optional) which clarifies what it means with the (out) annotation. https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/1551
2023-03-01 01:21:38 +01:00
* @cancellable: the #GCancellable to abort the shutdown.
* @callback: a #GAsyncReadyCallback to call when the request
libnm: add nm_client_wait_shutdown() function for cleaning up NMClient Add a fire-and-forget function to wait for shutdown to be complete. It's not entirely trivial to ensure all resources of NMClient are cleaned up. That matters only if NMClient uses a temporary GMainContext that the user wants to release while the application continues. For example, to do some short-lived operations an a worker thread. It's not trivial also because glib provides no convenient API to integrate a GMainContext in another GMainContext. We have that code as nm_utils_g_main_context_create_integrate_source(), so add a helper function to allow the user to do this. The function allows to omit the callback, in which case the caller wouldn't know when shutdown is complete. That would still be useful however, when integrating the client's context into the caller's context, so that the client's context gets automatically iterated until completion. The following test script will run out of file descriptors, when wait_shutdown() is not used: #!/bin/python import gi gi.require_version("NM", "1.0") from gi.repository import NM, GLib for i in range(1200): print(f">>>{i}") ctx = GLib.MainContext() ctx.push_thread_default() nmc = NM.Client.new() ctx.pop_thread_default() def cb(unused, result, i): try: NM.Client.wait_shutdown_finish(result) except Exception: # cannot happen assert False else: print(f">>>>> {i} complete") nmc.wait_shutdown(True, None, cb, i) while GLib.MainContext.default().iteration(False): pass
2022-10-05 12:22:51 +02:00
* is satisfied or %NULL if you don't care about the result of the
* method invocation.
* @user_data: the data to pass to @callback
*
* The way to stop #NMClient is by unrefing it. That will cancel all
* internally pending async operations. However, as async operations in
* NMClient use GTask, hence they cannot complete right away. Instead,
* their (internal) result callback still needs to be dispatched by iterating
* the client's main context.
*
* You thus cannot stop iterating the client's main context until
* everything is wrapped up. nm_client_get_context_busy_watcher()
* helps to watch how long that will be.
*
* This function automates that waiting. Like all glib async operations
* this honors the current g_main_context_get_thread_default().
*
* In any case, to complete the shutdown, nm_client_get_main_context()
* must be iterated. If the current g_main_context_get_thread_default() is
* the same as nm_client_get_main_context(), then @integrate_maincontext
* is ignored. In that case, the caller is required to iterate the context
* for shutdown to complete. Otherwise, if g_main_context_get_thread_default()
* differs from nm_client_get_main_context() and @integrate_maincontext
* is %FALSE, the caller must make sure that both contexts are iterated
* until completion. Otherwise, if @integrate_maincontext is %TRUE, then
* nm_client_get_main_context() will be integrated in g_main_context_get_thread_default().
* This means, the caller gives nm_client_get_main_context() up until the waiting
* completes, the function will acquire the context and hook it into
* g_main_context_get_thread_default().
* It is a bug to request @integrate_maincontext while having nm_client_get_main_context()
* acquired or iterated otherwise because a context can only be acquired once
* at a time.
*
* Shutdown can only complete after all references to @client were released.
*
* It is possible to call this function multiple times for the same client.
* But note that with @integrate_maincontext the client's context is acquired,
* which can only be done once at a time.
*
* It is permissible to start waiting before the objects is fully initialized.
*
* The function really allows two separate things. To get a notification (callback) when
* shutdown is complete, and to integrate the client's context in another context.
* The latter case is useful if the client has a separate context and you hand it
* over to another GMainContext to wrap up.
*
* The main use is to have a NMClient and a separate GMainContext on a worker
* thread. When being done, you can hand over the cleanup of the context
* to g_main_context_default(), assuming that the main thread iterates
* the default context. In that case, you don't need to care about passing
* a callback to know when shutdown completed.
*
* Since: 1.42
*/
void
nm_client_wait_shutdown(NMClient *client,
gboolean integrate_maincontext,
GCancellable *cancellable,
GAsyncReadyCallback callback,
gpointer user_data)
{
NMClientPrivate *priv;
WaitShutdownData *data;
gs_unref_object GTask *task = NULL;
GQuark quark = _wait_shutdown_get_quark();
GPtrArray *qdata_arr;
GSource *integration_source = NULL;
g_return_if_fail(NM_IS_CLIENT(client));
g_return_if_fail(!cancellable || G_IS_CANCELLABLE(cancellable));
priv = NM_CLIENT_GET_PRIVATE(client);
task = nm_g_task_new(NULL, cancellable, nm_client_wait_shutdown, callback, user_data);
if (integrate_maincontext && g_task_get_context(task) != priv->main_context) {
integration_source = nm_utils_g_main_context_create_integrate_source(priv->main_context);
g_return_if_fail(integration_source);
g_source_attach(integration_source, g_task_get_context(task));
}
data = g_slice_new(WaitShutdownData);
format: run nm-code-format Reformat with: clang-format version 19.1.0 (Fedora 19.1.0-1.fc41) https://gitlab.freedesktop.org/NetworkManager/NetworkManager/-/merge_requests/2046
2024-10-04 10:45:56 +02:00
*data = (WaitShutdownData) {
libnm: add nm_client_wait_shutdown() function for cleaning up NMClient Add a fire-and-forget function to wait for shutdown to be complete. It's not entirely trivial to ensure all resources of NMClient are cleaned up. That matters only if NMClient uses a temporary GMainContext that the user wants to release while the application continues. For example, to do some short-lived operations an a worker thread. It's not trivial also because glib provides no convenient API to integrate a GMainContext in another GMainContext. We have that code as nm_utils_g_main_context_create_integrate_source(), so add a helper function to allow the user to do this. The function allows to omit the callback, in which case the caller wouldn't know when shutdown is complete. That would still be useful however, when integrating the client's context into the caller's context, so that the client's context gets automatically iterated until completion. The following test script will run out of file descriptors, when wait_shutdown() is not used: #!/bin/python import gi gi.require_version("NM", "1.0") from gi.repository import NM, GLib for i in range(1200): print(f">>>{i}") ctx = GLib.MainContext() ctx.push_thread_default() nmc = NM.Client.new() ctx.pop_thread_default() def cb(unused, result, i): try: NM.Client.wait_shutdown_finish(result) except Exception: # cannot happen assert False else: print(f">>>>> {i} complete") nmc.wait_shutdown(True, None, cb, i) while GLib.MainContext.default().iteration(False): pass
2022-10-05 12:22:51 +02:00
.cancellable = nm_g_object_ref(cancellable),
.task = g_object_ref(task),
.result = -1,
.integration_source = integration_source,
.log_ptr = NM_HASH_OBFUSCATE_PTR(client),
};
/* The "data" itself stays alive as long as the task lives. That's important, because
* the callbacks _wait_shutdown_weak_ref_cb and _wait_shutdown_cancelled_cb
* rely on accessing "data", which must live long enough. */
g_task_set_task_data(task, data, _wait_shutdown_data_free);
g_weak_ref_init(&data->weak_ref, priv->context_busy_watcher);
NML_DBUS_LOG(_NML_NMCLIENT_LOG_LEVEL_COERCE(NML_DBUS_LOG_LEVEL_TRACE),
"nmclient[" NM_HASH_OBFUSCATE_PTR_FMT
"]: wait-shutdown (" NM_HASH_OBFUSCATE_PTR_FMT ")"
"%s",
data->log_ptr,
NM_HASH_OBFUSCATE_PTR(data),
integration_source ? " (integrated main source)" : "");
/* I don't think g_object_weak_ref() + GWeakRef can actually be used in
* a race-free way here, because g_object_weak_ref() has no GDestroyNotify
* and cannot keep task alive to avoid races. Instead, implement a weak pointer
* notification via the qdata.
*
* Yes, getting cancellation thread-safe is rather complicated here! I think
* the code is correct though, and you can cancel the operation from
* any thread without races. */
G_LOCK(wait_shutdown_mutex);
qdata_arr = g_object_get_qdata(priv->context_busy_watcher, quark);
if (!qdata_arr) {
qdata_arr = g_ptr_array_new();
g_object_set_qdata_full(priv->context_busy_watcher,
quark,
qdata_arr,
_wait_shutdown_qdata_cb);
}
/* data->task gets an additional reference, take it now. */
g_object_ref(data->task);
g_ptr_array_add(qdata_arr, data);
G_UNLOCK(wait_shutdown_mutex);
if (data->cancellable) {
/* Take an additional reference on task. */
data->cancellable_id = g_cancellable_connect(data->cancellable,
G_CALLBACK(_wait_shutdown_cancelled_cb),
g_object_ref(task),
g_object_unref);
}
}
/**
* nm_client_wait_shutdown_finish:
* @result: a #GAsyncResult obtained from the #GAsyncReadyCallback passed to nm_client_wait_shutdown()
* @error: return location for error or %NULL
*
* Returns: %TRUE if waiting is complete successfully. In that case, all resources of the
* nmclient are wrapped up and released. This can only fail by user cancellation.
*
libnm: fix a large amount of Since tags Some comments are malformed, some are missing altogether.
2022-11-07 21:13:28 +01:00
* Since: 1.42
libnm: add nm_client_wait_shutdown() function for cleaning up NMClient Add a fire-and-forget function to wait for shutdown to be complete. It's not entirely trivial to ensure all resources of NMClient are cleaned up. That matters only if NMClient uses a temporary GMainContext that the user wants to release while the application continues. For example, to do some short-lived operations an a worker thread. It's not trivial also because glib provides no convenient API to integrate a GMainContext in another GMainContext. We have that code as nm_utils_g_main_context_create_integrate_source(), so add a helper function to allow the user to do this. The function allows to omit the callback, in which case the caller wouldn't know when shutdown is complete. That would still be useful however, when integrating the client's context into the caller's context, so that the client's context gets automatically iterated until completion. The following test script will run out of file descriptors, when wait_shutdown() is not used: #!/bin/python import gi gi.require_version("NM", "1.0") from gi.repository import NM, GLib for i in range(1200): print(f">>>{i}") ctx = GLib.MainContext() ctx.push_thread_default() nmc = NM.Client.new() ctx.pop_thread_default() def cb(unused, result, i): try: NM.Client.wait_shutdown_finish(result) except Exception: # cannot happen assert False else: print(f">>>>> {i} complete") nmc.wait_shutdown(True, None, cb, i) while GLib.MainContext.default().iteration(False): pass
2022-10-05 12:22:51 +02:00
*/
gboolean
nm_client_wait_shutdown_finish(GAsyncResult *result, GError **error)
{
g_return_val_if_fail(nm_g_task_is_valid(result, NULL, nm_client_wait_shutdown), FALSE);
return g_task_propagate_boolean(G_TASK(result), error);
}
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
/*****************************************************************************
* Backported symbols. Usually, new API is only added in new major versions
src: drop most master references from the code While we cannot remove all the references to "master" we can remove most of them.
2024-05-27 17:58:08 +02:00
* of NetworkManager (that is, on "main" branch). Sometimes however, we might
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
* have to backport some API to an older stable branch. In that case, we backport
* the symbols with a different version corresponding to the minor API.
*
src: drop most master references from the code While we cannot remove all the references to "master" we can remove most of them.
2024-05-27 17:58:08 +02:00
* To allow upgrading from such a extended minor-release, "main" contains these
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
* backported symbols too.
*
* For example, 1.2.0 added nm_setting_connection_autoconnect_slaves_get_type.
* This was backported for 1.0.4 as nm_setting_connection_autoconnect_slaves_get_type@libnm_1_0_4
* To allow an application that was linked against 1.0.4 to seamlessly upgrade to
src: drop most master references from the code While we cannot remove all the references to "master" we can remove most of them.
2024-05-27 17:58:08 +02:00
* a newer major version, the same symbols is also exposed on "main". Note, that
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
* a user can only seamlessly upgrade to a newer major version, that is released
* *after* 1.0.4 is out. In this example, 1.2.0 was released after 1.4.0, and thus
* a 1.0.4 user can upgrade to 1.2.0 ABI.
*****************************************************************************/
NM_BACKPORT_SYMBOL(libnm_1_0_4,
NMSettingConnectionAutoconnectSlaves,
nm_setting_connection_get_autoconnect_slaves,
(NMSettingConnection * setting),
(setting));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_4,
GType,
nm_setting_connection_autoconnect_slaves_get_type,
(void),
());
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6,
NMMetered,
nm_setting_connection_get_metered,
(NMSettingConnection * setting),
(setting));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6, GType, nm_metered_get_type, (void), ());
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6,
NMSettingWiredWakeOnLan,
nm_setting_wired_get_wake_on_lan,
(NMSettingWired * setting),
(setting));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6,
const char *,
nm_setting_wired_get_wake_on_lan_password,
(NMSettingWired * setting),
(setting));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6, GType, nm_setting_wired_wake_on_lan_get_type, (void), ());
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6, const guint *, nm_utils_wifi_2ghz_freqs, (void), ());
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6, const guint *, nm_utils_wifi_5ghz_freqs, (void), ());
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6,
char *,
nm_utils_enum_to_str,
(GType type, int value),
(type, value));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: move backported symbols from libnm-core to libnm Backported symbols only make sense for libnm itself, not for libnm-core which is statically linked with NetworkManager and nm-ifcace-helper. Declaring the symbols in libnm-core, means that NetworkManager binary also contains them, although there are not used. Move them to libnm.
2016-10-12 11:26:26 +02:00
NM_BACKPORT_SYMBOL(libnm_1_0_6,
gboolean,
nm_utils_enum_from_str,
(GType type, const char *str, int *out_value, char **err_token),
(type, str, out_value, err_token));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
all: don't use gchar/gshort/gint/glong but C types We commonly don't use the glib typedefs for char/short/int/long, but their C types directly. $ git grep '\<g\(char\|short\|int\|long\|float\|double\)\>' | wc -l 587 $ git grep '\<\(char\|short\|int\|long\|float\|double\)\>' | wc -l 21114 One could argue that using the glib typedefs is preferable in public API (of our glib based libnm library) or where it clearly is related to glib, like during g_object_set (obj, PROPERTY, (gint) value, NULL); However, that argument does not seem strong, because in practice we don't follow that argument today, and seldomly use the glib typedefs. Also, the style guide for this would be hard to formalize, because "using them where clearly related to a glib" is a very loose suggestion. Also note that glib typedefs will always just be typedefs of the underlying C types. There is no danger of glib changing the meaning of these typedefs (because that would be a major API break of glib). A simple style guide is instead: don't use these typedefs. No manual actions, I only ran the bash script: FILES=($(git ls-files '*.[hc]')) sed -i \ -e 's/\<g\(char\|short\|int\|long\|float\|double\)\>\( [^ ]\)/\1\2/g' \ -e 's/\<g\(char\|short\|int\|long\|float\|double\)\> /\1 /g' \ -e 's/\<g\(char\|short\|int\|long\|float\|double\)\>/\1/g' \ "${FILES[@]}"
2018-07-11 07:40:19 +02:00
NM_BACKPORT_SYMBOL(libnm_1_2_4,
int,
nm_setting_ip_config_get_dns_priority,
(NMSettingIPConfig * setting),
(setting));
all: reformat all with new clang-format style Run: ./contrib/scripts/nm-code-format.sh -i ./contrib/scripts/nm-code-format.sh -i Yes, it needs to run twice because the first run doesn't yet produce the final result. Signed-off-by: Antonio Cardace <acardace@redhat.com>
2020-09-28 16:03:33 +02:00
libnm: add mdns backported symbols from 1.10.14 Add to master branch mdns symbols that were backported to 1.10.14 to allow seamless upgrading from 1.10 to 1.16.
2018-10-19 19:12:53 +02:00
NM_BACKPORT_SYMBOL(libnm_1_10_14,
NMSettingConnectionMdns,
nm_setting_connection_get_mdns,
(NMSettingConnection * setting),
(setting));
NM_BACKPORT_SYMBOL(libnm_1_10_14, GType, nm_setting_connection_mdns_get_type, (void), ());
libnm: add required-timeout backported symbol from 1.30.8 The nm_setting_ip_config_get_required_timeout() symbol was introduced in libnm 1.32.4 and then backported to 1.30.8. Export it also with version @libnm_1_30_8; this allows a program built against libnm 1.30.8 to keep working with later versions of the library.
2021-09-23 08:56:08 +02:00
NM_BACKPORT_SYMBOL(libnm_1_30_8,
int,
nm_setting_ip_config_get_required_timeout,
(NMSettingIPConfig * setting),
(setting));
libnm: add NMIPAddress and NMIPRoute dups backported symbols from 1.30.8 The nm_ip_address_dup() and nm_ip_route_dup() symbols were exposed in libnm 1.32 and then backported to 1.30.8. Export it also with version @libnm_1_30_8; this allows a program build against libnm 1.30.8 to keep working with later versions of the library. Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net>
2021-10-05 10:38:35 +02:00
NM_BACKPORT_SYMBOL(libnm_1_30_8,
NMIPAddress *,
nm_ip_address_dup,
(NMIPAddress * address),
(address));
NM_BACKPORT_SYMBOL(libnm_1_30_8, NMIPRoute *, nm_ip_route_dup, (NMIPRoute * route), (route));
libnm: add ethtool fec backported symbols from 1.50.4 Add to main branch ethtool fec symbols that were backported to 1.50.4 to allow seamless upgrading from 1.50 to 1.54.
2025-04-27 22:30:16 -04:00
libnm: add ethtool fec backported symbols from 1.46.8 Add to main branch ethtool fec symbols that were backported to 1.46.8 to allow seamless upgrading from 1.46 to 1.54.
2025-04-28 15:20:20 -04:00
NM_BACKPORT_SYMBOL(libnm_1_46_8,
gboolean,
nm_ethtool_optname_is_fec,
(const char *optname),
(optname));
NM_BACKPORT_SYMBOL(libnm_1_46_8, GType, nm_setting_ethtool_fec_mode_get_type, (void), ());
libnm: add ethtool fec backported symbols from 1.48.18 Add to main branch ethtool fec symbols that were backported to 1.48.18 to allow seamless upgrading from 1.48 to 1.54.
2025-04-28 15:07:49 -04:00
NM_BACKPORT_SYMBOL(libnm_1_48_18,
gboolean,
nm_ethtool_optname_is_fec,
(const char *optname),
(optname));
NM_BACKPORT_SYMBOL(libnm_1_48_18, GType, nm_setting_ethtool_fec_mode_get_type, (void), ());
libnm: add ethtool fec backported symbols from 1.50.4 Add to main branch ethtool fec symbols that were backported to 1.50.4 to allow seamless upgrading from 1.50 to 1.54.
2025-04-27 22:30:16 -04:00
NM_BACKPORT_SYMBOL(libnm_1_50_4,
gboolean,
nm_ethtool_optname_is_fec,
(const char *optname),
(optname));
NM_BACKPORT_SYMBOL(libnm_1_50_4, GType, nm_setting_ethtool_fec_mode_get_type, (void), ());
libnm: add safe file access backported symbols from 1.52.2 Add to branch nm-1-56 symbols for safe file access that were backported to 1.52.2 to allow seamless upgrading from 1.52 to 1.56.
2026-01-20 11:43:05 +01:00
NM_BACKPORT_SYMBOL(libnm_1_52_2,
char *,
nm_utils_copy_cert_as_user,
(const char *filename, const char *user, GError **error),
(filename, user, error));
NM_BACKPORT_SYMBOL(libnm_1_52_2,
gboolean,
nm_vpn_plugin_info_supports_safe_private_file_access,
(NMVpnPluginInfo * self),
(self));
libnm: add safe file access backported symbols from 1.54.3 Add to branch nm-1-56 symbols for safe file access that were backported to 1.54.3 to allow seamless upgrading from 1.54 to 1.56.
2026-01-20 11:57:30 +01:00
NM_BACKPORT_SYMBOL(libnm_1_54_3,
char *,
nm_utils_copy_cert_as_user,
(const char *filename, const char *user, GError **error),
(filename, user, error));
NM_BACKPORT_SYMBOL(libnm_1_54_3,
gboolean,
nm_vpn_plugin_info_supports_safe_private_file_access,
(NMVpnPluginInfo * self),
(self));
Reference in a new issue Copy permalink